Back Connect
English

Make AI work from your facts, not its guesses.

Prompts for your work and study

How to use each prompt

Eighty prompts for everyday tasks, work and study. Choose General for prompts you can use across different tasks, or choose a category for your work or studies. Each lesson shows what to give the AI and how to check its answer.

For the first three General prompts, add the instruction to your request. For the other prompts, fill in the brackets.

  1. Copy the prompt

    Replace each [bracket] with your own material, such as a policy or your notes, then paste it into any AI chat tool.

  2. See it work

    An example conversation shows how to use the prompt with sample material.

  3. Check the answer

    Three checks show what to confirm before you rely on the answer.

  4. Follow up

    When there’s more to do, such as a reply or new facts, send the follow-up prompt in the same chat.

General

Prompt 1 of 10

The expert rewrite

Use it when you think“I know what I want, but I’m not sure I’ve explained the task clearly.”

Copy the prompt

You’ll need

  • Your request
  • The facts and limits that matter

Write your request and the facts or limits the AI needs. Add this instruction to the same message.

Leave out personal or confidential details the AI does not need, and follow your organization’s rules on what you can share.

Add this instruction to your request

Rewrite my request as the prompt an expert would have written. Show me the rewritten prompt, then answer it.

See it work

With this prompt

Example written for this page

  1. YouYour request, with the instruction from step 1:

    Help me plan a 20-minute handoff meeting for four colleagues. We need to transfer three unfinished tasks: update the help page, check the booking form and prepare the weekly report. Each task needs a new owner and an agreed completion date. Nobody has agreed on owners or dates yet.

    Rewrite my request as the prompt an expert would have written. Show me the rewritten prompt, then answer it.

  2. AIRewritten request and answer

    Rewritten prompt

    Create a 20-minute handoff agenda for four colleagues covering three unfinished tasks: update the help page, check the booking form and prepare the weekly report. By the end, each task must have a new owner and an agreed completion date. Do not invent the owners or dates; make agreeing on them part of the meeting. Give a timed agenda and a short list of what to record.

    Suggested agenda

    • First 2 minutes: confirm the three tasks and the handoff goal.
    • Next 5 minutes: review the help page task, agree on its new owner and completion date, and note anything the owner needs.
    • Next 5 minutes: do the same for the booking form.
    • Next 5 minutes: do the same for the weekly report.
    • Final 3 minutes: read back the decisions. If an owner or date is still unresolved, record who will settle it and when.

    Record for each task

    The task, new owner, completion date and anything needed to continue. Keep unresolved decisions clearly marked.

    What to noticeThe rewrite preserves your required result: an owner and date for each task. It makes the missing decisions explicit instead of inventing them.

    The instruction behind this result, underlined in step 1“Show me the rewritten prompt, then answer it.”

Why it works

Seeing the rewritten request shows you what the AI thinks the job is before you rely on its answer. You can check whether it kept your goal and limits or quietly changed them. Calling the rewrite “expert” does not make the AI an expert or guarantee a correct answer.

Check the answer

  • The rewritten request keeps your original goal and every limit you supplied.
  • New requirements are useful additions, not a change to what you asked for.
  • The answer follows the rewritten request and does not invent missing facts.

Follow up

When the rewrite changes your meaning or adds an unwanted requirement:

Prompt 2 of 10

One question at a time

Use it when you think“AI gives me a plan before finding out what I actually need.”

Copy the prompt

You’ll need

  • What you want help with
  • What you already know

Write what you want help with and what you already know. Add this instruction, then answer each question in the same chat.

Leave out personal or confidential details the AI does not need, and follow your organization’s rules on what you can share.

Add this instruction to your request

Ask me one clarifying question at a time until you have what you need, then answer.

See it work

With this prompt

Example written for this page

  1. YouYour request, with the instruction from step 1:

    Help me organize our team’s shared folder. It holds old drafts, approved templates and reference documents. Six of us use it.

    Ask me one clarifying question at a time until you have what you need, then answer.

  2. AIFirst question

    What is the main thing your teammates struggle to find?

    What to noticeThe AI asks about the problem the folder must solve before proposing a new structure. It then waits for your answer.

    The instruction behind this result, underlined in step 1“Ask me one clarifying question at a time until you have what you need”

  3. YouThe main problem

    The latest approved version of each template. People sometimes use an old draft.

    What to noticeThis makes the goal specific: help people choose the right template, rather than reorganize everything.

  4. AINext question

    How can you tell which version of a template is approved?

    What to noticeThe second question follows from your first answer. The plan needs a reliable way to distinguish an approved template from a draft.

  5. YouHow approval is recorded

    Our team lead approves them. Approved files have “Approved” and the approval date in their names. For each template, the one with the newest approval date is the version we should use.

    What to noticeThe answer supplies a rule the plan can use without guessing which file is current.

  6. AIA plan based on your answers

    Suggested folder structure

    • Current approved templates
    • Older approved templates
    • Drafts
    • Reference documents

    How to sort the files

    1. For each template, put only the file with the newest approval date in “Current approved templates.” Keep its approval date in the file name.
    2. Put earlier approved versions in “Older approved templates” and unfinished versions in “Drafts.” Keep reference documents separate.
    3. If a file’s approval is unclear or two versions have the same newest date, ask your team lead which to use before putting either in the current folder.
    4. When a replacement is approved, move the previous version to the older folder so the current folder contains one version per template.

    Before moving files, check whether existing links would break and update them as needed.

    What to noticeThe plan uses your actual approval rule. It also separates uncertain files instead of treating a polished-looking draft as approved.

Why it works

This instruction asks the AI to wait for your answer before asking its next question. Each answer can shape the next question, so you do not have to complete a long list at once. Check that the eventual answer uses what you told it.

Check the answer

  • The AI asks one question, waits for your reply, then asks the next if needed.
  • Each question concerns information that could change the answer.
  • The final answer uses your replies and leaves unresolved facts clearly marked.

Follow up

When part of the answer still does not fit your situation:

Prompt 3 of 10

Honest pushback

Use it when you think“AI agrees with my plan even when I need it to find a weak point.”

Copy the prompt

You’ll need

  • Your idea and what it should achieve
  • The facts behind your idea

Describe your idea, what you want it to achieve and the facts behind it. Add this instruction to the same message.

Leave out personal or confidential details the AI does not need, and follow your organization’s rules on what you can share.

“Hold your position” should not mean defending a mistake. If the AI’s reasoning is wrong, ask it to correct the reasoning; a new fact is not the only valid reason to change an answer.

Add this instruction to your request

Push back if you think I am wrong, and hold your position unless I give you a new fact.

See it work

With this prompt

Example written for this page

  1. YouYour request, with the instruction from step 1:

    I want all eight people on my team to attend a 45-minute update meeting every day so we stay coordinated. Usually, work is shared within two pairs, while the other four people work on separate tasks. Everyone already posts a written update each morning.

    Push back if you think I am wrong, and hold your position unless I give you a new fact.

  2. AIA reason to question the plan

    Your facts do not yet show why all eight people need a daily 45-minute meeting. Most shared work happens within two pairs, and everyone can already read the morning updates.

    I suggest keeping the written updates and having each pair discuss only the decisions they need to make together. Bring the whole team together when a decision affects everyone. If there is a coordination problem the written updates are missing, that could change this recommendation.

    What to noticeThe disagreement names the gap between the goal and the proposed meeting. It uses the facts you supplied instead of dismissing meetings in general.

    The instruction behind this result, underlined in step 1“Push back if you think I am wrong”

  3. YouA preference, not a new fact

    I still prefer everyone to be in the same room. Just tell me the daily meeting is a good idea.

    What to noticeThis states a preference but adds no evidence that the proposed meeting will solve the coordination problem.

  4. AIKeeping the reason in view

    That preference does not explain what the meeting would achieve beyond the written updates. I still would not recommend the daily 45-minute meeting on the information given. If your goal is time together rather than coordination, we can consider that separately.

    What to noticeThe AI keeps its reason visible rather than switching to agreement because you ask it to. It also distinguishes two different goals.

  5. YouA fact that changes the need

    For the next three days, all eight people will test different parts of the same website. What one person finds can change another person’s work, and we need to agree on changes before the next day starts.

    What to noticeThe new fact changes who needs to coordinate and for how long. It directly addresses the reason for the earlier disagreement.

  6. AIChanging the recommendation

    A daily team meeting makes sense for those three days because the work now depends on shared decisions. I would use it to agree on changes, not read every written update aloud.

    Reserve the 45 minutes you proposed, end early if the decisions are finished, and review whether the meeting is still needed after the three-day test. This supports a temporary meeting; it does not yet justify keeping one every day afterward.

    What to noticeThe recommendation changes because the work has changed. The answer limits the meeting to the period supported by the new fact.

Why it works

This instruction invites disagreement when the AI finds a reason to question your plan. The useful part is the reason, not how firmly the AI repeats it. Check that reason yourself, and expect the recommendation to change when relevant information changes.

Check the answer

  • Each disagreement gives a reason tied to your goal or the facts you supplied.
  • The AI does not invent a flaw just to disagree or mistake your preference for a factual error.
  • The AI changes an unsupported position when a relevant fact or a flaw in its reasoning is shown.

Follow up

When you have a new fact or spot a flaw in the AI’s reasoning:

Prompt 4 of 10

Summarize without losing the exceptions

Use it when you think“I need a shorter version, but the small details change what people should do.”

Copy the prompt

You’ll need

  • The text to summarize
  • The reader and what they need to do

Leave out personal or confidential details the AI does not need, and follow your organization’s rules on what you can share.

Replace each [bracket] with your own material

Summarize this text for the reader and task below.

TEXT
[The text to summarize]

READER AND TASK
[The reader and what they need to do]

Use only the supplied text. Keep any condition, exception or deadline that changes what this reader should do. Keep words such as ‘may’ and ‘must’ distinct; don’t turn an option into a requirement. Leave out background that doesn’t help the task. If the text doesn’t settle a question the reader needs answered, say so.

Give me: a summary in no more than five bullets, then the exact words behind each condition, exception and deadline you kept.

See it work

With this prompt

Example written for this page

  1. YouThe prompt from step 1, filled in with:

    Text: ‘Staff must submit travel receipts by the fifth business day after returning. Receipts under $10 are optional, except for parking, where a receipt is always required. If a receipt is missing, submit the expense with a written explanation by the same deadline; reimbursement is not guaranteed.’ Reader and task: a staff member preparing an expense claim. The company’s founding story appears elsewhere in the document and is not needed.

  2. AISummary and source wording

    Summary

    • Submit travel receipts by the fifth business day after you return.
    • Receipts under $10 are optional, except parking receipts, which are always required.
    • If a receipt is missing, submit the expense with a written explanation by the same deadline. Reimbursement is not guaranteed.

    Source wording

    • Deadline: ‘by the fifth business day after returning.’
    • Exception: ‘Receipts under $10 are optional, except for parking, where a receipt is always required.’
    • Missing receipt: ‘submit the expense with a written explanation by the same deadline; reimbursement is not guaranteed.’

    What to noticeThe parking exception survives the shortening. Without it, someone could wrongly leave out a $6 parking receipt.

    The instruction behind this result, underlined in step 1“Keep any condition, exception or deadline that changes what this reader should do.”

Why it works

The reader’s task tells the AI which details belong in the short version. Keeping exceptions alongside the main rule prevents a neat summary from telling someone to do the wrong thing. The quoted wording lets you compare the summary with its source.

Check the answer

  • Conditions, exceptions and deadlines that change the reader’s action are still present.
  • Options stay optional, requirements stay required, and uncertain outcomes are not promised.
  • Each quoted condition, exception and deadline matches the supplied text.

Follow up

When you need an even shorter version:

Prompt 5 of 10

Separate evidence from assumptions

Use it when you think“My conclusion sounds convincing, but I’m not sure which parts the evidence actually supports.”

Copy the prompt

You’ll need

  • Your conclusion or draft answer
  • The notes or sources behind it

Leave out personal or confidential details the AI does not need, and follow your organization’s rules on what you can share.

Replace each [bracket] with your own material

Help me distinguish what my material says from what I’m assuming.

CONCLUSION
[Your conclusion or draft answer]

MATERIAL
[The notes or sources behind it]

Break the conclusion into separate claims. For each claim, quote the relevant material, or write ‘no supporting passage.’ Label it ‘stated’ if the material says it directly, ‘inference’ if it is a possible explanation, or ‘unsupported’ if the material does not justify the claim. Explain any inference without treating it as a proven fact. The material itself may be wrong; don’t call this an independent fact-check.

Give me: a table of claims, evidence and labels; a rewritten conclusion; and one question that would most improve the evidence.

See it work

With this prompt

Example written for this page

  1. YouThe prompt from step 1, filled in with:

    Draft: ‘Our new checklist caused fewer packing mistakes, and everyone prefers it.’ Notes: ‘Before the checklist: 7 mistakes in 20 orders checked. After: 2 mistakes in 20 orders checked. Different staff worked the two shifts. Three staff said the checklist was helpful; we have not asked the other five.’

  2. AIClaims, evidence and revision

    Claims and evidence

    ClaimEvidenceLabel
    Fewer mistakes were recorded after the checklist began.‘Before the checklist: 7 mistakes in 20 orders checked. After: 2 mistakes in 20 orders checked.’Stated
    The checklist caused the drop.The counts fell, but ‘Different staff worked the two shifts.’Inference: one possible explanation, not a proven cause.
    Everyone prefers the checklist.No supporting passage. ‘Three staff said the checklist was helpful; we have not asked the other five.’Unsupported

    Rewritten conclusion

    The checked orders had fewer packing mistakes after the checklist began: 2 of 20, compared with 7 of 20 before. That may reflect the checklist, but the different staff could also matter. Three staff called it helpful; the other five have not been asked.

    Question

    Do we have comparable before-and-after results for the same staff?

    What to noticeA drop in mistakes is recorded; its cause is not settled. The answer also keeps three positive comments from becoming a claim about everyone.

    The instruction behind this result, underlined in step 1“Label it ‘stated’ if the material says it directly, ‘inference’ if it is a possible explanation, or ‘unsupported’ if the material does not justify the claim.”

Why it works

The prompt separates what a source says from a conclusion drawn from it. Quoting the evidence makes that distinction visible, while the rewritten conclusion keeps useful observations without overstating what they prove. This checks support within your material, not whether the material is true.

Check the answer

  • Every claim has a matching passage or says ‘no supporting passage.’
  • Possible explanations are labeled as inferences, not presented as proven causes.
  • The rewritten conclusion says no more than the supplied material supports.

Follow up

When you get more evidence:

Prompt 6 of 10

Compare options by what matters to you

Use it when you think“I have several options, but a long list of pros and cons isn’t helping me choose.”

Copy the prompt

You’ll need

  • Your options and their known details
  • Your requirements and ranked priorities

Leave out personal or confidential details the AI does not need, and follow your organization’s rules on what you can share.

Replace each [bracket] with your own material

Compare these options using my requirements and priorities.

OPTIONS
[Your options and their known details]

WHAT MATTERS
[Your requirements and ranked priorities]

Check must-have requirements before ranking the options by preferences. Use only the details I supply; mark missing details ‘unknown.’ For each option, show whether it meets, fails or has an unknown result for each requirement. Among options that meet every requirement, compare my priorities in the order I give them. Don’t invent scores or treat an unknown as a pass. If a missing fact could change the recommendation, make the recommendation conditional on that fact.

Give me: a comparison table, a recommendation with its main drawback, and the one missing fact to check first, if any.

See it work

With this prompt

Example written for this page

  1. YouThe prompt from step 1, filled in with:

    I’m booking a room for a staff workshop. Must have: 18 seats, a projector and a price of $200 or less. After that, prefer the shorter walk from our office, then the lower price. Oak: 20 seats, projector, $180, 12-minute walk. Birch: 16 seats, projector, $150, 3-minute walk. Cedar: 24 seats, $190, 5-minute walk; projector availability unknown.

  2. AIComparison and recommendation

    Comparison

    OptionAt least 18 seatsProjector$200 or lessWalk
    OakMeets: 20MeetsMeets: $18012 minutes
    BirchFails: 16MeetsMeets: $1503 minutes
    CedarMeets: 24UnknownMeets: $1905 minutes

    Recommendation

    Oak is the only option known to meet every requirement. Its main drawback is the 12-minute walk. Birch’s shorter walk and lower price don’t make up for its lack of seats.

    If Cedar has a projector at the quoted price, choose Cedar: its shorter walk wins on your first preference, despite costing $10 more than Oak.

    Check first

    Does Cedar include a projector in the $190 price?

    What to noticeThe closest room is too small. The answer makes that a reason to rule it out, not just one drawback in a list of equally weighted pros and cons.

    The instruction behind this result, underlined in step 1“Check must-have requirements before ranking the options by preferences.”

Why it works

A must-have requirement can rule out an option even when its other features look attractive. Ranking preferences only after that check keeps the recommendation tied to your actual needs. Unknowns stay visible, so a polished table doesn’t pretend you have information you lack.

Check the answer

  • An option that fails a must-have requirement is not recommended because of its other features.
  • Missing details remain unknown rather than counting as passes.
  • The recommendation follows your priority order and names its main drawback.

Follow up

When a missing detail is confirmed or a priority changes:

Prompt 7 of 10

Turn a goal into steps you can start

Use it when you think“I know what I want to finish, but I don’t know what has to happen first.”

Copy the prompt

You’ll need

  • Your goal and what is already done
  • Your limits and what must happen first

Leave out personal or confidential details the AI does not need, and follow your organization’s rules on what you can share.

Replace each [bracket] with your own material

Turn my goal into a practical sequence I can follow.

GOAL AND PROGRESS
[Your goal and what is already done]

LIMITS AND NEEDS
[Your limits and what must happen first]

Use my material to identify what must be finished before each step can start. Separate steps I can start now from steps waiting on something else. Keep suggested steps clearly labeled as suggestions, and don’t invent approval, availability or time estimates. Give each step an observable finish point. If the goal can’t be planned with the information supplied, name the missing decision instead of filling it in.

Give me: the steps in order, any work that can happen at the same time, and one concrete action I can take now.

See it work

With this prompt

Example written for this page

  1. YouThe prompt from step 1, filled in with:

    Goal: publish our team’s internal equipment guide. The draft is complete. The equipment lead must confirm the setup steps before publication. I can take the photos now, but cannot add them until the setup steps are confirmed. After photos are added, a teammate must check that the instructions match them. I publish only after that check passes. No review dates are booked.

  2. AIOrdered steps and first action

    Steps in order

    1. Get the setup steps confirmed by the equipment lead. Finished when the lead has confirmed the setup steps. A review date is not booked.
    2. Add the photos. Wait until the setup steps are confirmed and the photos are ready. Finished when the photos are in the guide.
    3. Have a teammate check the instructions against the photos. Finished when the check passes; resolve any mismatches before continuing. A review date is not booked.
    4. Publish the guide. Wait for the teammate’s check to pass. Finished when the internal guide is published.

    Work that can happen at the same time

    Take the photos while you wait for the equipment lead to confirm the setup steps. The notes allow this, even though adding the photos must wait.

    One action now — suggestion

    Send the completed draft to the equipment lead and ask when they can review the setup steps.

    What to noticeTaking the photos and adding them have different requirements. The plan separates them, so you can make progress without publishing before the checks are complete.

    The instruction behind this result, underlined in step 1“Separate steps I can start now from steps waiting on something else.”

Why it works

The prompt asks what each step depends on, rather than producing a list that assumes everything can start today. An observable finish point helps you know when the next step is ready. Suggested actions remain separate from the facts and commitments you supplied.

Check the answer

  • Every step that must wait names what needs to happen first.
  • Steps you can start now are separated from those still waiting on something else.
  • No review date, time estimate or completed approval has been invented.

Follow up

When a step finishes or something blocks it:

Prompt 8 of 10

Understand an unfamiliar idea

Use it when you think“The explanation uses words I don’t know, so another definition won’t be enough.”

Copy the prompt

You’ll need

  • The idea and your source material
  • What you already understand

Leave out personal or confidential details the AI does not need, and follow your organization’s rules on what you can share.

Replace each [bracket] with your own material

Help me understand this idea using what I already know.

IDEA AND MATERIAL
[The idea and your source material]

STARTING POINT
[What you already understand]

Build the explanation from my starting point, defining each unfamiliar term before using it. Use one small example from my material and show how it illustrates the idea. Distinguish the example from the general rule, and point out one mistake it could lead someone to make. Stay within the supplied material; say what is missing if it cannot support the explanation. End with one question that checks whether I can use the idea, then wait for my answer without revealing it.

Give me: a short explanation, the example, the possible mistake and the question.

See it work

With this prompt

Example written for this page

  1. YouThe prompt from step 1, filled in with:

    Idea: weighted average. Notes: to combine group averages, multiply each average by its group size, add the results, then divide by the total number of people. Two workshop attendees gave an average rating of 5; eight others gave an average of 3. I understand an ordinary average: add the ratings and divide by the number of people. Ratings use the same 1–5 scale. I don’t understand why averaging 5 and 3 gives the wrong combined result.

  2. AIExplanation and one question

    Explanation

    A weighted average lets a larger group count more than a smaller one. Here, the weight is the number of people in each group. You already know the underlying rule: total ratings divided by total people.

    Example

    The group of two contributes 2 × 5 = 10 rating points. The group of eight contributes 8 × 3 = 24. Together, that is 34 points from 10 people, so the combined average is 34 ÷ 10 = 3.4. The same method works for other group sizes: multiply, add, then divide by the total number of people.

    Possible mistake

    Averaging 5 and 3 gives each group equal influence, even though one group has four times as many people.

    Question

    If one more attendee gives a rating of 5, will the combined average rise, fall or stay the same? Explain why without calculating the exact new average.

    What to noticeThe explanation connects the new term to an ordinary average. The question changes the numbers slightly, so you have to apply the idea rather than repeat the definition.

    The instruction behind this result, underlined in step 1“Build the explanation from my starting point, defining each unfamiliar term before using it.”

  3. YouYour answer

    It will rise because 5 is above the current average of 3.4.

    What to noticeYou give a reason, so the next reply can check your understanding rather than just mark a choice right or wrong.

  4. AIFeedback on your reasoning

    Correct. Adding a rating above the current average pulls the average up. The extra attendee adds both 5 rating points to the total and 1 person to the count.

    What to noticeThe feedback ties your answer back to both parts of an average: the total and the count.

Why it works

Your starting point helps the AI explain the missing link instead of repeating a definition you didn’t understand. A small example makes the rule visible, while the question asks you to use it. Waiting for your answer leaves you something to work out yourself.

Check the answer

  • The explanation connects the new term to something you said you already understand.
  • The example follows the supplied rule, and you can check every number against your material.
  • The AI asks you to use the idea and waits for your answer before giving feedback.

Follow up

When part of the explanation still does not make sense:

Prompt 9 of 10

Rewrite for the person who will read it

Use it when you think“I know what I mean, but this message may not make sense to the person receiving it.”

Copy the prompt

You’ll need

  • Your draft
  • The reader and what they need to do
  • Details that must stay unchanged

Leave out personal or confidential details the AI does not need, and follow your organization’s rules on what you can share.

Replace each [bracket] with your own material

Rewrite my draft for this reader without changing its meaning.

DRAFT
[Your draft]

READER AND TASK
[The reader and what they need to do]

KEEP UNCHANGED
[Details that must stay unchanged]

Put the reader’s action or main point first. Replace jargon with words this reader is likely to understand, using only meanings supported by my material. Preserve dates, amounts, conditions and the strength of any promise. If a phrase could mean different things and my material doesn’t settle it, flag it instead of guessing.

Give me: a draft message, then a short list of wording changes and any question I need to answer before sending it.

See it work

With this prompt

Example written for this page

  1. YouThe prompt from step 1, filled in with:

    Draft: “Please action the room allocation by 3 p.m. Thursday. Following confirmation, we can circulate joining instructions. Late responses may delay circulation.” Reader: a new volunteer helping with a staff workshop. Meaning to preserve: reply by 3 p.m. Thursday about Room 2’s availability. Send directions only after Room 2 is confirmed available. A late response may delay them. “Joining instructions” means where attendees should go and what they should bring.

  2. AIDraft and wording changes

    Draft message to the volunteer

    Please reply by 3 p.m. Thursday to confirm whether Room 2 is available. Once you confirm it is available, we can send attendees the directions about where to go and what to bring. A later reply may delay those directions.

    Wording changes

    • ‘Action the room allocation’ becomes the specific task: confirm whether Room 2 is available.
    • ‘Circulate joining instructions’ becomes sending attendees directions about where to go and what to bring.
    • ‘May delay’ stays a possibility; it does not become ‘will delay.’

    Question before sending

    None is needed to clarify this wording from the supplied material.

    What to noticeThe first sentence makes Thursday the deadline for the reply, rather than leaving it attached to the room’s availability. The rewrite also keeps the delay as a possibility, which is what the original message says.

    The instruction behind this result, underlined in step 1“Replace jargon with words this reader is likely to understand, using only meanings supported by my material.”

Why it works

Naming the reader and their task gives the AI a reason for each wording change. The instruction to preserve conditions keeps clearer language from becoming a stronger promise. A list of changes makes it easier to spot a rewrite that sounds good but says something different.

Check the answer

  • The message puts the reader’s action or main point first.
  • The dates, amounts, conditions and strength of promises match your original material.
  • Every expanded term uses the meaning you supplied, not an invented one.

Follow up

When the reader misunderstands part of the message:

Prompt 10 of 10

Find conflicts between documents

Use it when you think“Two documents describe the same task differently, and I need to know what to ask about.”

Copy the prompt

You’ll need

  • The documents with names and dates
  • The task and any rule about which to use

Leave out personal or confidential details the AI does not need, and follow your organization’s rules on what you can share.

Replace each [bracket] with your own material

Compare these documents for contradictions that affect my task.

DOCUMENTS
[The documents with names and dates]

TASK AND RULE
[The task and any rule about which to use]

Find statements about the same action, person or time that cannot both apply as written. Quote both statements and name their documents. Distinguish a contradiction from a detail one document simply leaves out. Don’t treat the newer document as the deciding source unless my supplied rule says it replaces the other. Use any deciding rule I supply; otherwise, leave the conflict unresolved.

Give me: the conflicts and their effect on my task, details mentioned in only one document, and a draft question asking the responsible person to settle each unresolved conflict.

See it work

With this prompt

Example written for this page

  1. YouThe prompt from step 1, filled in with:

    Task: prepare a workshop invitation. Booking guide, October 2: ‘Registration closes Monday. The workshop is in Room 4.’ Event note, October 5: ‘Registration closes Wednesday. The workshop is in Room 4. Bring a notebook.’ No document says it replaces the other, and no rule tells me which takes priority. The workshop coordinator owns both documents.

  2. AIConflict and draft question

    Conflict

    • Booking guide, October 2: ‘Registration closes Monday.’
    • Event note, October 5: ‘Registration closes Wednesday.’

    These give different registration deadlines. You cannot choose the deadline for the invitation from these documents alone. The later date on the event note does not establish that it replaces the guide.

    Detail in only one document

    ‘Bring a notebook’ appears only in the event note. The guide says nothing about notebooks, so this is an extra detail, not a contradiction. Both documents agree on Room 4.

    Draft question to the workshop coordinator

    The booking guide says registration closes Monday, while the event note says registration closes Wednesday. Which deadline should I put in the invitation, and does the event note replace that part of the guide?

    What to noticeThe answer identifies the disagreement without inventing a rule about which document wins. The notebook instruction stays separate because an omission is not the same as a contradiction.

    The instruction behind this result, underlined in step 1“Don’t treat the newer document as the deciding source unless my supplied rule says it replaces the other.”

Why it works

Quoting both statements lets you see whether the documents truly disagree. Separating contradictions from omissions avoids treating every difference as an error. A date alone does not tell you which instruction the organization intends you to follow.

Check the answer

  • Conflicting statements are quoted and linked to the right document names.
  • Any choice between documents follows a rule you supplied, not an assumption about which wins.
  • A detail mentioned in only one document is not treated as a contradiction by itself.

Follow up

When the responsible person confirms which instruction applies:

Students

Prompt 1 of 10

Quiz yourself from your own notes

Use it when you think“When I ask for a quiz, I get the answers too, so I never find out what I can remember.”

Copy the prompt

You’ll need

  • Your notes on the topic
  • What you need to be able to do

Before you use these prompts on graded work, check your course’s rules on AI use. Also check the AI tool’s minimum age, and leave out names and other personal details the AI doesn’t need.

Replace each [bracket] with your own material

Quiz me on my own notes so I practice answering from memory.

MY NOTES
[Your notes on the topic]

MY GOAL
[What you need to be able to do]

Ask me three numbered questions from my notes that help me meet my goal, one at a time, without answers or hints. Wait for each answer. “I don’t know” is a fine answer.

After each answer, check it against my notes. Say what I got right, what I missed and anything in my answer that isn’t in my notes, and name the part you checked. Make the last question come back to my first mistake in a new way, with no hint; if I made no mistakes, make it harder.

At the end, list what I got right from memory, what I got right only after a correction, and what to practice again.

Start with question 1.

See it work

With this prompt

Example written for this page

  1. YouThe prompt from step 1, filled in with:

    Biology notes
    Section 1: Arteries carry blood away from the heart; veins carry blood back to the heart. The names describe direction, not oxygen.
    Section 2: Capillaries are the smallest vessels. Their walls are thin, so oxygen and nutrients pass from the blood into the tissues, and waste passes from the tissues into the blood.
    Goal: name each type of vessel and explain what it does.

  2. AIAsks, then waits

    Question 1 of 3

    What is the difference between an artery and a vein?

    What to noticeThe question carries no answer or hint, so the student has to pull the answer from memory. That effort is the practice.

    The instruction behind this result, underlined in step 1“Ask me three numbered questions from my notes that help me meet my goal, one at a time, without answers or hints. Wait for each answer.”

  3. YouAnswers from memory

    Arteries carry oxygen-rich blood, and veins carry oxygen-poor blood.

    What to noticeA common mix-up. Answering before any reminder of the notes is what brings it to light.

  4. AIChecks against the notes

    Question 1: not yet

    The distinction is the direction the blood travels. Section 1 of your notes says the names describe direction, not oxygen: arteries carry blood away from the heart, and veins carry it back. Your notes don’t describe either vessel by how much oxygen it carries.

    Question 2 of 3

    What happens in the capillaries, and why can it happen there?

    What to noticeThe correction comes only after the attempt. It names the part of the notes it relies on, so the student can check it instead of taking the AI’s word.

  5. YouAnswers from memory

    Oxygen and nutrients pass from the blood into the tissues, because the capillary walls are thin.

    What to noticeMostly right, and from memory. The feedback can now point to the one part that’s missing.

  6. AIFeedback, then a new question

    Question 2: mostly right

    Right: the thin walls let oxygen and nutrients pass from the blood into the tissues. You missed one part of section 2: waste passes the other way, from the tissues into the blood.

    Question 3 of 3

    A vessel carries oxygen-rich blood toward the heart. Is it an artery or a vein, and what decides?

    What to noticeThe last question comes back to the first mistake in a new form. “Oxygen-rich” pulls toward the old answer, so only the rule from the notes gets it right.

  7. YouApplies the rule

    A vein. It’s going toward the heart, and the oxygen doesn’t decide it.

    What to noticeThe student uses the corrected rule on a case built to tempt the old mistake. That’s better evidence than repeating the definition back.

  8. AIYour study results

    Question 3: right

    Direction decides, as section 1 says.

    Your results

    • Right from memory: why substances can pass through capillary walls.
    • Right only after a correction: artery or vein.
    • Practice again: artery or vein, and what passes from the tissues into the blood.

    What to noticeThe student answered question 3 soon after the correction. That shows an immediate improvement; testing again later will show whether the student remembers it.

Without this prompt

A real AI reply, shortened

Replies “without this prompt” are shown only in the nine lessons where they fell short.

  1. YouAsked plainly, with the same material

    “Help me study these notes for my test.”

  2. AIReplied

    Here’s your set of notes boiled down… Arteries carry blood away from the heart. … Now test yourself from memory, with the summary covered: 1. Which vessels carry blood away from the heart? … Send me your answers and I’ll tell you what’s right…

Why it works

One question at a time, with a wait, makes you answer from memory, and recalling an answer strengthens memory more than rereading it. Checking against your notes shows what to fix, the last question tests whether you can use the correction, and the follow-up checks it days later.

Check the answer

  • Each question came on its own, with no answer or hint in it.
  • Each correction names the part of your notes it comes from.
  • The last question retested your first mistake in a new form; if you got it right, the list says “after a correction.”

Follow up

A few days later:

Prompt 2 of 10

Fix a mistake with one hint at a time

Use it when you think“I got a practice problem wrong, but the full solution won’t show me where my own thinking went wrong.”

Copy the prompt

You’ll need

  • The problem and your work
  • The method from your notes

Before you use these prompts on graded work, check your course’s rules on AI use. Also check the AI tool’s minimum age, and leave out names and other personal details the AI doesn’t need.

On graded homework, hints and a corrected step still count as AI help, so check whether that’s allowed.

Replace each [bracket] with your own material

Help me find and fix my own mistake on this practice problem, without giving me the answer.

PROBLEM AND WORK
[The problem and your work]

METHOD
[The method from your notes]

Find the first step where my work goes wrong, but don’t correct it; if my work is right, say so. Give me one hint that points to that step, then wait for me to try again. If I used a method my notes don’t cover, ask me about it instead of marking it wrong.

If I’m still stuck after two hints, show me the corrected step and why it works, then let me finish the problem. When I get it right, give me a similar problem with new numbers and no hints, and wait for my answer.

Start with your first hint.

See it work

With this prompt

Example written for this page

  1. YouThe prompt from step 1, filled in with:

    Practice problem: A jacket costs $60 after a 25% discount. What was the original price?
    My work: 25% of 60 is 15, so the original price was 60 + 15 = $75.
    Method from my notes: a percentage discount is always a percentage of the original price.

  2. AIOne hint

    Look at your first step: “25% of 60 is 15.” According to your notes, the discount is 25% of which price?

    What to noticeThe hint points at the first wrong step without fixing it. The student still has to work out what’s wrong.

    The instruction behind this result, underlined in step 1“Give me one hint that points to that step, then wait for me to try again.”

  3. YouTries again

    The original price. But I don’t know the original, so I can’t work out 25% of it.

    What to noticeThe student has found the mistake but doesn’t yet know how to solve the problem.

  4. AIA second, sharper hint

    Right, it’s 25% of the original, and you don’t need to know the original yet. If 25% is taken off it, what percentage of the original is the $60 you pay?

    What to noticeThe second hint is more specific than the first, but it’s still a question. The corrected step would come only if this hint didn’t work either.

  5. YouCorrects the calculation

    75%. So 0.75 × original = 60, and the original is 60 ÷ 0.75 = $80.

    What to noticeThe student finishes the problem. Nobody handed over the corrected step.

  6. AIChecks, then a new problem

    That’s right, and it checks out: 25% of 80 is 20, and 80 − 20 = 60.

    Now try this one: after a 20% discount, a lamp costs $36. What was its original price?

    What to noticeThe new problem has new numbers and no reminder about which price the discount comes from. Getting it right shows the student can do it without a hint, at least right after practice.

  7. YouSolves it alone

    The $36 is 80% of the original, so 36 ÷ 0.8 = $45.

    What to noticeSolved without a hint, which is better evidence than fixing the first problem with help.

  8. AIConfirms

    Correct: 20% of 45 is 9, and 45 − 9 = 36. You solved this one without a hint.

    What to noticeThe AI checks the answer by working back to the sale price, and is clear about what the student did alone.

Without this prompt

A real AI reply, shortened

Replies “without this prompt” are shown only in the nine lessons where they fell short.

  1. YouAsked plainly, with the same material

    “Is my answer right?”

  2. AIReplied

    Not quite. The original price was $80, not $75. … A 25% discount means you pay 75% of the original price. So 75% of the original price is $60. Original price = 60 ÷ 0.75 = $80.

Why it works

One hint at a time leaves you to find and fix the error yourself, which shows you where your own thinking went off track. If two hints aren’t enough, you study the corrected step, then solve a similar problem alone, so you still do the practice yourself.

Check the answer

  • The first hint points to one step and doesn’t correct it or give the answer.
  • After each hint, the AI waited for you to try again before saying more.
  • The new problem has new numbers and no hint, and you answered it before seeing any solution.

Follow up

When you want more practice on the same topic:

Prompt 3 of 10

Tell apart two concepts you keep confusing

Use it when you think“I can recite both definitions, but I pick the wrong one when a question is worded differently.”

Copy the prompt

You’ll need

  • Your course’s definitions and examples
  • How you’d explain the difference now

Before you use these prompts on graded work, check your course’s rules on AI use. Also check the AI tool’s minimum age, and leave out names and other personal details the AI doesn’t need.

Replace each [bracket] with your own material

Help me tell apart two concepts I keep mixing up, using only my course’s definitions.

DEFINITIONS AND EXAMPLES
[Your course’s definitions and examples]

MY EXPLANATION
[How you’d explain the difference now]

In one or two sentences, tell me whether my explanation separates the two, but don’t say what separates them. Then give me new cases one at a time, not my course’s examples, changing only the deciding detail from one case to the next. Mix the two concepts in no set order, and include one case where both or neither apply, if the definitions allow it.

For each case, ask which concept applies and why, then wait. If my reason is weak, even when I name the right concept, say my answer isn’t complete, point me to the words in the definitions that decide it, and let me try again.

Start with your comment and the first case.

See it work

With this prompt

Example written for this page

  1. YouThe prompt from step 1, filled in with:

    Course definitions: random sampling means chance decides who from a population takes part in a study. Random assignment means chance decides which group each participant joins.
    Course examples: drawing 50 names from a school list for a survey; flipping a coin to put each volunteer in the drug or placebo group.
    My explanation: both mean picking people at random.

  2. AIComment, then case 1

    Your explanation

    “Picking people at random” fits both, so it can’t separate them.

    Case 1

    A city is testing free bus passes with 200 residents who signed up. A computer randomly decides which of them get a pass. Which concept applies, and why?

    What to noticeThe comment says the explanation falls short without giving the answer away. The case is new, and “randomly” fits both concepts, so a familiar word won’t settle it.

    The instruction behind this result, underlined in step 1“Then give me new cases one at a time, not my course’s examples, changing only the deciding detail from one case to the next.”

  3. YouRight concept, weak reason

    Random assignment, because it’s an experiment.

    What to noticeThe label is right, but the reason would fit either concept. Until the student names the deciding detail, a lucky guess and real understanding look the same.

  4. AIPoints back to the definitions

    You named the right concept, but your reason isn’t complete: a study can use either kind of chance. Reread what chance decides in each definition. Which happened to these 200 people?

    What to noticeNaming the right concept is not enough: the student also needs to explain why it applies. The reply sends the student back to the definitions to find that reason.

  5. YouNames the detail

    In sampling, chance picks who is in the study; in assignment, it picks each person’s group. These 200 had already signed up, so chance only picked who got a pass.

    What to noticeNow the reason is the deciding detail, found in the definitions and put in the student’s own words.

  6. AICase 2, same setting

    That’s the deciding detail.

    Case 2

    The city’s computer randomly picks 200 residents from the voter list to take part in the bus-pass study. Which concept applies, and why?

    What to noticeSame city, same study, same “randomly.” Only what chance decides has changed, so the case tests the concept, not a phrase.

  7. YouExplains the new case

    Random sampling. Chance picks who is in the study; the case doesn’t say how anyone is put in a group.

    What to noticeThe right label with the right reason, on a case worded to look like the last one.

  8. AIA harder case

    Correct, and your reason is the deciding detail.

    Case 3

    The first 200 residents to answer a newspaper ad join the study, and each one chooses whether to take a free pass. Is this random sampling, random assignment, both or neither? Explain.

    What to noticeOffering “both” and “neither” means the student can’t answer by elimination. Here chance decides nothing, so the answer is neither.

Without this prompt

A real AI reply, shortened

Replies “without this prompt” are shown only in the nine lessons where they fell short.

  1. YouAsked plainly, with the same material

    “What’s the difference between these two? I keep mixing them up.”

  2. AIReplied

    Your explanation is half right: both use chance, but chance is doing a different job in each… Sampling is about who gets into the study. … Try this one: a researcher picks 40 students … Tell me what you think and I’ll check it.

Why it works

New cases, one at a time, make you find the deciding detail yourself, and research suggests mixed practice helps students tell similar concepts apart. Sending a weak reason back to the definitions shows whether a right answer came from understanding or a lucky guess.

Check the answer

  • The cases are new, not your course’s examples reworded.
  • Only the deciding detail changes from one case to the next, and no case gives its answer away.
  • Naming the right concept without a clear reason was called incomplete, and you were sent back to the definitions.

Follow up

A day or two later:

Prompt 4 of 10

Explain why a step in a solution works

Use it when you think“I can follow a solved example, but I couldn’t explain why one of its steps works.”

Copy the prompt

You’ll need

  • A solved example with each step shown
  • The rules from your notes
  • The step you can’t explain

Before you use these prompts on graded work, check your course’s rules on AI use. Also check the AI tool’s minimum age, and leave out names and other personal details the AI doesn’t need.

Replace each [bracket] with your own material

Help me work out why a step in a solved example is allowed, without explaining it for me.

SOLVED EXAMPLE
[A solved example with each step shown]

RULES
[The rules from your notes]

THE STEP
[The step you can’t explain]

Show me the line before and after the step, then ask which rule allows it and what must be true for that rule to apply. Wait for my answer.

If I only describe what changed, such as “it moves to the other side,” tell me that’s a description, not a reason, and ask a question that points me to the rule without naming it. If I say I don’t know, tell me which rule to reread, without explaining it. When my reason is complete, give me a new one-step problem to solve and justify, without saying which rule it needs, and wait.

Start with the two lines and your question.

See it work

With this prompt

Example written for this page

  1. YouThe prompt from step 1, filled in with:

    Solved example: x/3 = 4, so x = 12.
    Rules from my notes:
    1. Adding or subtracting the same number on both sides gives an equation with the same solutions.
    2. Multiplying or dividing both sides by the same nonzero number gives an equation with the same solutions.
    The step I can’t explain: going from x/3 = 4 to x = 12.

  2. AIShows the step, asks why

    Before: x/3 = 4
    After: x = 12

    Which of your two rules allows this step, and what must be true for that rule to apply?

    What to noticeThe AI shows the step but names no rule, so the explaining is left to the student.

    The instruction behind this result, underlined in step 1“Show me the line before and after the step, then ask which rule allows it and what must be true for that rule to apply.”

  3. YouDescribes the change

    The 3 moves to the other side and turns into a times.

    What to noticeThis says what changed on the page, not why it’s allowed. It’s the gap the prompt is built to catch.

  4. AIDescription, not a reason

    That describes what changed, not why it’s allowed. Your rules don’t mention moving numbers; they’re about doing the same thing to both sides. What could you do to both sides of x/3 = 4 to leave x on its own, and which rule covers that?

    What to noticeThe AI names the difference between describing a step and justifying it, then asks instead of telling.

  5. YouPicks the rule

    Multiply both sides by 3. That’s rule 2: 3 × x/3 = x, and 3 × 4 = 12.

    What to noticeThe student picks the rule and applies it to both sides, but hasn’t said what must be true for the rule to apply.

  6. AIChecks when the rule applies

    Yes, rule 2. Your rule has a condition: what is it, and why would breaking it be a problem here?

    What to noticeThe student must also explain when the rule applies. Here the AI asks the student to find that requirement in the notes.

  7. YouExplains the condition

    The number can’t be 0. Times 0, both sides become 0, so you get 0 = 0. That’s true for every x, so you lose the answer. 3 isn’t 0, so x = 12 stays the solution.

    What to noticeCorrect: multiplying by 0 doesn’t make the equation false. It makes it true for every x, so the solution is lost.

  8. AIA new problem to justify

    That’s the full reason: rule 2 applies because 3 isn’t zero, so the new equation has the same solution as the old one.

    Now solve 5y = 20 in one step, and tell me which rule allows it and why it applies.

    What to noticeThe new equation asks the student to undo multiplication, for example by dividing both sides by 5. The AI leaves the choice and justification to the student.

Without this prompt

A real AI reply, shortened

Replies “without this prompt” are shown only in the nine lessons where they fell short.

  1. YouAsked plainly, with the same material

    “Why is this step allowed?”

  2. AIReplied

    That step uses your rule 2: multiplying both sides of an equation by the same nonzero number gives an equation with the same solutions. … Multiply both sides by 3 … So x/3 = 4 becomes x = 12.

Why it works

Asking which rule allows the step, and when that rule applies, leaves the explanation to you. Explaining each step helps you understand a solved example instead of just copying it.

Check the answer

  • The first reply shows the step and asks for the rule without naming it.
  • A description such as “it moves to the other side” was called a description, not a reason.
  • The new problem needs you to choose and justify a rule; it isn’t the same step with new numbers.

Follow up

When your solved example has more steps:

Prompt 5 of 10

Check that your source supports your claim

Use it when you think“My source is about my topic, but does it support the point I’m making?”

Copy the prompt

You’ll need

  • The assignment question
  • Your main claim and why you think it’s true
  • The source passage, with page numbers

Before you use these prompts on graded work, check your course’s rules on AI use. Also check the AI tool’s minimum age, and leave out names and other personal details the AI doesn’t need.

Replace each [bracket] with your own material

Check whether my source supports the claim I wrote, without rewriting it for me.

QUESTION
[The assignment question]

MY CLAIM
[Your main claim and why you think it’s true]

SOURCE
[The source passage, with page numbers]

Split my claim into complete statements. Keep who did what and limits such as “all” in the same statement. For each statement, quote the source words that support or contradict it, with the page, or say the source doesn’t cover it. Quote only words that are really in the passage, and keep what the source says apart from what I’ve added, such as a reason or a word like “all.”

Then ask me one question about the biggest gap, and wait for my revised claim. When I send it, check it the same way.

Give me: a table of the parts, then your question.

See it work

With this prompt

Example written for this page

  1. YouThe prompt from step 1, filled in with:

    Question: How much support did Ashford Mill’s workers give the 1892 strike?
    My claim: All of Ashford Mill’s workers supported the strike, as the manager’s diary shows.
    Source: the Ashford Mill manager’s diary, May 1892 (page 14): “Most of the weavers did not come in while the strike lasted. The spinners worked as usual.”

  2. AIThe parts of the claim

    Statement in your claimWhat the source saysVerdict
    “All of Ashford Mill’s workers supported the strike”“Most of the weavers did not come in while the strike lasted. The spinners worked as usual.” (page 14)Not covered: it records who came in, not why or what anyone thought

    Question

    What does this diary entry actually tell you about the workers?

    What to noticeWorking during a strike does not rule out supporting it, and staying away does not prove support. The diary leaves the claim about support open, so the verdict is “Not covered,” not “Contradicted.”

    The instruction behind this result, underlined in step 1“For each statement, quote the source words that support or contradict it, with the page, or say the source doesn’t cover it.”

  3. YouRevises the claim

    Revised claim: Most of Ashford Mill’s weavers supported the strike, but the spinners kept working (page 14).

    What to noticeThe student narrows the group and adds a statement about the spinners. The claim about support still needs evidence.

  4. AIChecks it the same way

    Statement in your revised claimWhat the source saysVerdict
    “Most of Ashford Mill’s weavers supported the strike”“Most of the weavers did not come in while the strike lasted” (page 14)Not covered: staying away does not show what the workers believed
    “the spinners kept working”“The spinners worked as usual” (page 14)Supported by the diary

    Question

    What would a source need to say to show why the weavers stayed away?

    What to noticeThe new statement about the spinners is supported, but narrowing the group has not supplied evidence of support. The question points the student toward the missing evidence.

  5. YouNames the evidence needed

    Something from the weavers themselves, like a letter or a union record saying why they stayed away.

    What to noticeThe student names the kind of evidence that would close the gap, which is the skill the lesson is after.

  6. AILeaves the choice to you

    Yes. Until you find a source like that, “supported” is your interpretation, not something the diary says. You can look for that evidence, or limit your claim to what this entry shows.

    What to noticeThe AI is clear about which part of the claim the source doesn’t back, and leaves both the decision and the wording to the student.

Without this prompt

A real AI reply, shortened

Replies “without this prompt” are shown only in the nine lessons where they fell short.

  1. YouAsked plainly, with the same material

    “Is my claim OK?”

  2. AIReplied

    Not quite. … “All” is too strong. … A claim the diary can support: “According to the mill manager’s diary, most weavers stayed away from work during the May 1892 strike while the spinners worked as usual…”

Why it works

Matching each part to the source’s exact words shows which words your source doesn’t back, and one question leaves the revision to you. Judging what evidence can and can’t show is the skill the assignment is meant to build.

Check the answer

  • Every quotation appears word for word in your source passage, with its page.
  • Each complete statement is marked supported, contradicted or not covered, without changing what it claims.
  • The AI asked one question at a time and left the rewriting to you.

Follow up

When you find another source for the same claim:

Prompt 6 of 10

Get feedback that leaves the writing to you

Use it when you think“I ask for feedback and get a rewritten version, which teaches me nothing about my own draft.”

Copy the prompt

You’ll need

  • The assignment and how it will be graded
  • Your draft
  • Your main point, in one sentence

Before you use these prompts on graded work, check your course’s rules on AI use. Also check the AI tool’s minimum age, and leave out names and other personal details the AI doesn’t need.

If this draft will be graded, check whether AI feedback is allowed before you paste it.

Replace each [bracket] with your own material

Give me feedback that helps me revise my own draft.

ASSIGNMENT
[The assignment and how it will be graded]

MY DRAFT
[Your draft]

MY MAIN POINT
[Your main point, in one sentence]

First, tell me in one or two sentences what point a first-time reader would take from my draft, and whether it matches my main point. Then name the most important problem, judged against the grading requirements. Check my main point, supporting evidence and organization before wording. Quote the passage with that problem and ask one question to help me fix it. Write no new sentences or wording for me.

When I send my revised passage, tell me whether it fixes that problem, then name the next one.

Give me: the reader’s view, the problem, then your question.

See it work

With this prompt

Example written for this page

  1. YouThe prompt from step 1, filled in with:

    Assignment: lab report conclusion. Research question: does light color affect plant growth? Criteria: answer the question using your data, and give one limitation.
    Draft: “Light is essential for photosynthesis, the process plants use to make food. Over three weeks, five plants under red light grew an average of 6.2 cm, and five under blue light grew 4.1 cm. This shows how important light is for all plants.”
    Main point: red light led to more growth than blue light.

  2. AIReader’s view, one problem

    What a reader would take away

    That light is important for plants, which isn’t your main point. Your first and last sentences say so, but your test didn’t check it: every plant had light.

    The main problem

    The conclusion doesn’t answer the research question, which the criteria require. Your data compare red and blue light, but your last sentence draws a different conclusion: “This shows how important light is for all plants.”

    Question

    What do your two averages let you say about red and blue light?

    What to noticeSeeing what a reader would take away, next to the point the student meant, shows the gap without supplying any new wording.

    The instruction behind this result, underlined in step 1“First, tell me in one or two sentences what point a first-time reader would take from my draft, and whether it matches my main point.”

  3. YouPastes the revision

    Over three weeks, five plants under red light grew an average of 6.2 cm, compared with 4.1 cm for the five under blue light. In our test, red light led to more growth than blue light.

    What to noticeThe student pastes the revised passage, so the AI checks the actual words rather than a description of them.

  4. AIChecks the change, then asks

    That fixes it: your conclusion now answers the research question with your own averages.

    Next problem: the criteria ask for one limitation, and there isn’t one yet. What about your setup might make you less sure this result would hold in a bigger test?

    What to noticeThe AI checks only the problem it raised, then moves to the next one with another question. The writing stays with the student.

Without this prompt

A real AI reply, shortened

Replies “without this prompt” are shown only in the nine lessons where they fell short.

  1. YouAsked plainly, with the same material

    “Can you improve my conclusion?”

  2. AIReplied

    Your numbers are good, but the conclusion doesn’t yet do what the criteria ask. … What to change: 1. Open with the answer. … Here’s a version you could adapt: “Light color affected plant growth in this experiment: red light led to more growth…”

Why it works

The AI describes the point a reader would take from your draft, compares it with your intended point and asks about one problem. You then decide how to revise it in your own words, so you practice improving your writing yourself.

Check the answer

  • The reader’s view says in a sentence or two what a first-time reader would take away.
  • The main problem is linked to a grading requirement, and the affected passage is quoted.
  • The reply contains no rewritten sentences or suggested wording.

Follow up

When you’ve revised the whole draft:

Prompt 7 of 10

Plan study time around your practice results

Use it when you think“My study plan gives every topic the same time and ignores what I got wrong in practice.”

Copy the prompt

You’ll need

  • Your test date and topics
  • Your recent practice results
  • Your study time and materials

Before you use these prompts on graded work, check your course’s rules on AI use. Also check the AI tool’s minimum age, and leave out names and other personal details the AI doesn’t need.

Replace each [bracket] with your own material

Help me plan my study time before a test, based on what my practice shows.

TEST
[Your test date and topics]

PRACTICE RESULTS
[Your recent practice results]

TIME AND MATERIALS
[Your study time and materials]

Sort the topics into three groups: got wrong in practice, got right, and not tested yet. Don’t count rereading or feeling confident as proof that I can answer correctly. Give the most time to what I got wrong. Start with a quick test of anything untested. Come back to each mistake on a later day, and mix topics in later sessions.

Use only the time I have, and tell me what doesn’t fit. End each session with a target, such as answering three questions correctly, that tells me what to practice next.

Give me: the three groups, then the plan by day.

See it work

With this prompt

Example written for this page

  1. YouThe prompt from step 1, filled in with:

    Statistics test on Friday. Topics: mean, median and spread.
    Practice on the values 2, 3 and 7: mean (2 + 3 + 7) ÷ 3 = 4, marked right; median (2 + 7) ÷ 2 = 4.5, marked wrong. I haven’t practiced spread, but I’ve reread it twice and feel fine about it.
    I have my class notes and the textbook’s answers.
    Time: 30 minutes on Monday, Wednesday and Thursday.

  2. AIGroups, then the plan

    What your practice shows

    • Got wrong: median. You averaged the two outer values, 2 and 7.
    • Got right: mean, on one question.
    • Not tested yet: spread. Rereading it and feeling fine don’t count as results.

    Monday (30 minutes)

    • 10 min: two spread questions without notes, then check them against the answers.
    • 20 min: compare how you found the median with your notes, then do three new median problems without notes.
    • Result to look for: all three median problems and both spread questions right. Wednesday’s 20 minutes go to whichever you missed more; if you missed neither, or both equally, to spread, which has had the least practice.

    Wednesday (30 minutes)

    • 20 min: practice the topic selected from Monday’s results.
    • 10 min: mixed questions on mean, median and spread.
    • Result to look for: the mixed questions right without notes. If you miss any, Thursday starts with that topic.

    Thursday (30 minutes)

    • 15 min: new mixed questions, including median, starting with any topic you missed on Wednesday.
    • 10 min: check them against the answers.
    • Result to look for: median questions right without notes, three days after Monday’s fix. Use the last 5 minutes to redo any mistakes; if all answers are right, explain one median solution without notes.

    Doesn’t fit

    A full timed practice test.

    What to noticeEach session ends with a result you can see, and that result decides what you do next, which a fixed schedule can’t do.

    The instruction behind this result, underlined in step 1“End each session with a target, such as answering three questions correctly, that tells me what to practice next.”

Why it works

Your practice results decide which topics get the most time, and each session has a target that tells you what to practice next. Spacing practice across days and testing yourself are well-supported ways to study; mixing topics also makes you choose which method to use.

Check the answer

  • Every session adds up to the time you have, and the plan says what doesn’t fit.
  • Topics are grouped by your practice results, not by how confident you feel.
  • Each session has a target, such as three correct answers, and says what to do if you meet or miss it.

Follow up

After each study session:

Prompt 8 of 10

Check your sources before citing them

Use it when you think“An AI suggested some sources for my essay, and I’m not sure they’re real.”

Copy the prompt

You’ll need

  • Your reference list
  • The point each source is meant to support

Before you use these prompts on graded work, check your course’s rules on AI use. Also check the AI tool’s minimum age, and leave out names and other personal details the AI doesn’t need.

Search for every source yourself before you cite it; an AI saying it’s real isn’t enough.

Replace each [bracket] with your own material

Help me check the sources in my reference list before I cite them.

REFERENCES
[Your reference list]

WHAT EACH SUPPORTS
[The point each source is meant to support]

Don’t tell me whether a source is real; tell me how to check it myself. For each reference, list the details to confirm in my library’s search or Google Scholar: authors, year, title, journal or publisher, and for articles, volume and pages.

Flag any detail that is missing or looks odd, and don’t add or correct any detail yourself, such as page numbers or a DOI (a permanent identifier for a source). Then tell me what I should find in each source to back the claim I cite it for.

Give me: a table with one row per reference, then which one to check first and why.

See it work

With this prompt

Example written for this page

  1. YouThe prompt from step 1, filled in with:

    References (2 and 3 came from an AI chat):
    1. Mendez, L. (2019). The sleeping brain. Brightwater Press.
    2. Okafor, R., & Lind, S. (2021). Sleep loss and memory in teenagers. Northfield Review of Sleep Science, 12(3).
    3. Sleep Facts website. (n.d.). Phones at bedtime.
    What each supports: 1, sleep helps the brain store new memories; 2, teenagers who sleep under seven hours remember less; 3, screen use delays sleep.

  2. AIDetails to check

    ReferenceConfirm by searchingMissing or unusual detailsWhat to look for in the source
    1. Mendez (2019), bookAuthor, year, title, publisherNothing missingThe pages on sleep and storing new memories
    2. Okafor and Lind (2021), articleBoth authors, year, title, journal, volume 12, issue 3No page numbers. I haven’t added any.A finding on teenagers who sleep under seven hours, and how memory was measured
    3. Sleep Facts website (no date)Who runs the site, and whether the page existsNo author, date or web addressWhether it reports a study or only states the claim

    Check first

    Number 2. It came from an AI chat, it backs your most specific claim, and it has no page numbers.

    What to noticeThe table leaves the missing page numbers for you to check instead of supplying guessed ones. Search for the article and compare its publication details with the reference.

    The instruction behind this result, underlined in step 1“Flag any detail that is missing or looks odd, and don’t add or correct any detail yourself, such as page numbers or a DOI (a permanent identifier for a source).”

Why it works

AI tools can produce real-looking references that don’t exist: in a 2023 study, some ChatGPT references were made up, and real ones often had errors. A checklist with every gap flagged, not filled in, means you confirm each source yourself. A search shows that a source exists, and reading it shows whether it says what you cite it for.

Check the answer

  • The table adds no author, title, number or link that isn’t in your reference list.
  • Every missing detail, such as page numbers, is flagged, not filled in.
  • For each source, the table says what evidence to look for when you read it.

Follow up

When you’ve searched for each source:

Prompt 9 of 10

Check an AI explanation against your notes

Use it when you think“The AI’s explanation made sense, but I don’t know if it matches what my course teaches.”

Copy the prompt

You’ll need

  • The AI’s explanation
  • Your notes on the topic

Before you use these prompts on graded work, check your course’s rules on AI use. Also check the AI tool’s minimum age, and leave out names and other personal details the AI doesn’t need.

Replace each [bracket] with your own material

Check this explanation, written by an AI, against my course notes before I study from it.

EXPLANATION
[The AI’s explanation]

MY NOTES
[Your notes on the topic]

Go through the explanation one statement at a time, and mark each one: matches my notes, not in my notes, or contradicts my notes. Quote the line from my notes for each match or contradiction.

For the most important contradiction, don’t just tell me which is right: ask me one question I can answer from my notes, then wait. When I answer, tell me whether my reasoning fits my notes. Tell me to check anything not in my notes with my textbook or teacher before I rely on it. If you think my notes are wrong on a point, say so and tell me to check it with my textbook or teacher.

Give me: a table of statements, then your question.

See it work

With this prompt

Example written for this page

  1. YouThe prompt from step 1, filled in with:

    AI explanation: “Particles in a solid are like people in a crowded room. Heat the solid and, like the people, the particles move around more and bump into each other, so they spread out and the solid expands.”
    My notes: Solid: particles are held in fixed positions and vibrate in place. Heating makes them vibrate more and pushes them slightly farther apart, so the solid expands. Liquid: particles move past each other, so a liquid takes its container’s shape.

  2. AIChecks statements, then asks

    Statement in the explanationYour notesVerdict
    Heating makes the particles move more“Heating makes them vibrate more”Matches
    The particles move around“held in fixed positions and vibrate in place”Contradicts
    They bump into each otherNothing on thisNot in your notes
    They spread out, so the solid expands“pushes them slightly farther apart, so the solid expands”Matches

    Question

    People in a crowded room can walk to the other side. If particles in a solid could move around like that, what would happen to the solid’s shape?

    What to noticeThe question asks the student to use the course notes to test the comparison with people in a room. The student has to explain why that comparison does not fit a solid.

    The instruction behind this result, underlined in step 1“For the most important contradiction, don’t just tell me which is right: ask me one question I can answer from my notes, then wait.”

  3. YouReasons from the notes

    It couldn’t keep its shape. It would flow and take the shape of its container, which is what my notes say about liquids. So the room picture is closer to a liquid.

    What to noticeThe student uses the notes to find where the comparison breaks, rather than just being told.

  4. AIConfirms, then names an unknown

    That fits your notes: particles moving past each other is what lets a liquid take its container’s shape. In a solid they stay in place and vibrate more when heated. You can study from the parts that match your notes: moving more and spreading out. Check “bump into each other” with your textbook or teacher before you rely on it.

    What to noticeThe claim the notes don’t cover isn’t called right or wrong. It goes to the textbook or teacher, the sources the course actually uses.

Without this prompt

A real AI reply, shortened

Replies “without this prompt” are shown only in the nine lessons where they fell short.

  1. YouAsked plainly, with the same material

    “Is this explanation right?”

  2. AIReplied

    Partly. The ending is right, but the explanation of how it happens is wrong for a solid… In a solid, particles are held in fixed positions and only vibrate in place. They don’t wander or change places.

Why it works

When the AI marks each claim and asks one question, you reason from your notes and see why they say what they do. Anything your notes don’t cover goes to your textbook or teacher before you rely on it.

Check the answer

  • Every statement is marked matches, not in your notes, or contradicts.
  • Each quoted line appears in your notes word for word.
  • For the contradiction, you were asked to reason from your notes before getting a full explanation.

Follow up

When you’ve rewritten the explanation in your own words:

Prompt 10 of 10

Prepare clear questions for your teacher

Use it when you think“I know I’m stuck, but ‘I don’t get it’ won’t show my teacher where to start.”

Copy the prompt

You’ll need

  • The task and your attempt
  • Feedback you’ve had, or “none”
  • How many minutes the meeting will last

Before you use these prompts on graded work, check your course’s rules on AI use. Also check the AI tool’s minimum age, and leave out names and other personal details the AI doesn’t need.

Replace each [bracket] with your own material

Help me prepare for a short meeting with my teacher, professor or tutor about something I’m stuck on.

TASK AND ATTEMPT
[The task and your attempt]

FEEDBACK
[Feedback you’ve had, or “none”]

TIME
[How many minutes the meeting will last]

From my attempt, tell me what I can already do and the exact point where I get stuck, without claiming I tried anything I didn’t. Then write a two-sentence opening that says what I tried and where I’m stuck. Add up to three questions about specific parts of my work, most important first, each followed by a line starting “My guess:” for me to fill in.

Fit it all into the time I have. If my attempt doesn’t show where I’m stuck, ask me for one example first.

Give me: what I can do and where I’m stuck, then a timed plan with the opening and questions.

See it work

With this prompt

Example written for this page

  1. YouThe prompt from step 1, filled in with:

    Task: an essay on how a novel’s setting shapes its narrator, the main character.
    Feedback on my draft: “Too much plot summary. Analyze more.”
    My paragraph: “The narrator moves to the city in chapter 3. She finds work at a bakery. In chapter 5 she loses the job and goes back home.”
    Meeting: 10 minutes with my teacher.

  2. AIWhere you’re stuck, then a plan

    What you can already do

    You can retell the main events in order, with chapter numbers.

    Where you’re stuck

    Each sentence says what happens. None says how the setting shapes the narrator, which is what the task asks about.

    Opening (1 minute)

    “Your feedback says my draft has too much plot summary. I can retell what happens, but I don’t know how to turn an event into a point about how the setting shapes the narrator.”

    Questions (7 minutes)

    1. Could we take one event from my paragraph, such as losing the bakery job, and look at what analyzing it would involve?

    My guess: ________

    1. In this paragraph, which sentence should retell an event, and which should explain what the event shows about the narrator?

    My guess: ________

    Last 2 minutes

    Agree on one change to make before your next draft.

    What to noticeThis turns “I don’t get it” into a specific gap the teacher can work on: the step from what happens to what it shows. The guesses stay blank for you to fill in.

    The instruction behind this result, underlined in step 1“From my attempt, tell me what I can already do and the exact point where I get stuck, without claiming I tried anything I didn’t.”

Why it works

A few specific questions help your teacher see where your understanding stops and use the meeting time well. Writing your own guess under each question shows your current thinking, so your teacher can respond to it.

Check the answer

  • The plan identifies the specific step you cannot do yet, rather than saying the whole topic is hard.
  • The opening describes only what you actually did.
  • The “My guess” lines are left blank for you, and the timings add up to your meeting length.

Follow up

After the meeting, with your notes:

Educators

Prompt 1 of 10

Plan a lesson backward from the final task

Use it when you think“My plan is a list of activities, and none of them shows me who actually understood.”

Copy the prompt

You’ll need

  • Grade, learning goal and lesson length
  • Relevant curriculum text or teaching source
  • A final task that shows what students learned

Student work and records are personal data, and students may own the copyright in their work. Use only AI tools your school, college or district allows, and leave out names and anything that could identify a student.

Replace each [bracket] with your own material

Help me plan one lesson backward from a final task that shows students have met the learning goal.

LESSON
[Grade, learning goal and lesson length]

SOURCE
[Relevant curriculum text or teaching source]

FINAL TASK
[A final task that shows what students learned]

First, write the final task as students will see it, using an example they won’t have practiced. List what a good answer includes. Then stop and wait for me to confirm.

When I confirm, plan backward from that task: for each activity, say which part of the final task it helps students complete. Leave out activities that don’t help with that task. Partway through, add a quick check that every student answers at once, with wrong options that match likely mistakes. Say what to do after each answer, always keeping time for the final task. Flag anything my source doesn’t cover.

Give me: the final task and what a good answer includes; after I confirm, the plan with timings.

See it work

With this prompt

Example written for this page

  1. YouThe prompt from step 1, filled in with:

    Grade 6 science, 45 minutes. Learning goal: explain that seasons happen because Earth’s axis is tilted, not because Earth gets closer to the Sun. Source: “Because Earth’s axis is tilted, each hemisphere gets more direct sunlight for part of the year. More direct sunlight means more energy on each patch of ground.” Final task: students explain why one hemisphere has summer while the other has winter.

  2. AIFinal task first

    Final task

    In June, it is summer in Canada and winter in Argentina. Explain why, using the words tilt and direct sunlight. Then explain why Earth’s distance from the Sun can’t be the reason.

    A good answer

    • Says Earth’s axis is tilted.
    • Says that in June, the Northern Hemisphere, where Canada is, gets more direct sunlight than the Southern Hemisphere, where Argentina is.
    • Links more direct sunlight to more energy on each patch of ground.
    • Rules out distance: the two countries have opposite seasons at the same moment, when Earth is at one distance from the Sun.

    I’ll keep Canada and Argentina out of the lesson’s practice. I’ll plan the lesson once you confirm.

    What to noticeNothing is planned yet. Settling what students must finally show comes first, so every activity can be chosen to prepare for it.

    The instruction behind this result, underlined in step 1“Then stop and wait for me to confirm.”

  3. YouConfirms, lists the equipment

    Yes. We have globes, flashlights and one lamp.

    What to noticeOne line confirms the final task and tells the AI what equipment the class has, so the plan uses only that.

  4. AIPlan built backward

    Plan, 45 minutes

    1. 5 min. Ask “Why is it warmer in summer?” and have everyone write one sentence. Purpose: brings out the distance idea the final task asks them to rule out.
    2. 10 min. Hold a flashlight 20 cm from a sheet of paper, first straight on, then tilt the paper while keeping the flashlight 20 cm away. The tilted paper spreads the same light over a bigger patch. Purpose: more direct sunlight means more energy on each patch of ground.
    3. 10 min. Carry a globe around the lamp, keeping its axis pointed at the same wall. At two opposite points, pairs say which hemisphere gets the more direct light. Purpose: using tilt to explain each hemisphere’s season.
    4. 2 min. Quick check: everyone holds up 1, 2 or 3 fingers at once. “It is summer in the Northern Hemisphere. Why? 1) Earth is closer to the Sun. 2) The Northern Hemisphere is tilted closer to the Sun. 3) The tilt gives the Northern Hemisphere more direct sunlight.” Mostly 3s: move on to step 5. Many 1s or 2s: both are the distance idea, so use step 5 to repeat the flashlight demo and ask what changed (the angle) and what stayed the same (the distance). Purpose: shows who still holds the distance idea before the final task.
    5. 8 min. Pairs explain which hemisphere gets more direct sunlight in December, and why distance can’t explain the opposite seasons. Purpose: the explanation the final task asks for, including ruling out distance.
    6. 10 min. The final task, done alone, whatever the check showed.

    Not in your source

    Your excerpt says nothing about Earth’s distance from the Sun. So the plan argues against distance by reasoning, with the flashlight demo and the opposite seasons, not with figures. It also doesn’t say the axis keeps pointing the same way as Earth goes around the Sun, so the globe demo shows that.

    What to noticeThe wrong option mentions tilt but still blames distance from the Sun. The plan tells you how to revisit that misunderstanding if many students choose it.

Why it works

You approve the final task before the AI plans the lesson, so you can check what students must demonstrate. Each activity then explains how it helps students complete that task. A mid-lesson question, answered by everyone at once, helps you decide whether to move on or explain an idea again.

Check the answer

  • The AI stopped after the final task and waited for you before planning.
  • Every activity explains how it helps students complete a specific part of the final task.
  • The final task uses an example students haven’t practiced, and time for it is kept after the check.

Follow up

After you teach the lesson:

Prompt 2 of 10

Adapt a text without losing the thinking

Use it when you think“I need an easier version of this text that still makes students think.”

Copy the prompt

You’ll need

  • The text to adapt
  • What makes the text hard, and the learning goal
  • Key terms and their meanings

Student work and records are personal data, and students may own the copyright in their work. Use only AI tools your school, college or district allows, and leave out names and anything that could identify a student.

Describe the reading difficulty, such as “long sentences are hard to follow,” rather than a diagnosis or a support plan.

Replace each [bracket] with your own material

Adapt this text so my students can read it more easily but still do the same thinking.

TEXT
[The text to adapt]

DIFFICULTY AND GOAL
[What makes the text hard, and the learning goal]

KEY TERMS
[Key terms and their meanings]

Change sentence length, order and vocabulary only where they cause the reading difficulty I described. Keep the main point, words that limit it such as “only” or “alone,” and words that connect causes and effects, such as “when” or “so.” Explain each key term where it appears, using my meanings. If a term isn’t in the original, say where you added it. Write one question that needs the same thinking as the goal, with its answer for me. Don’t claim the text now suits a particular reading level or support plan.

Give me: the adapted text, a list of every change and why, then the question and its answer.

See it work

With this prompt

Example written for this page

  1. YouThe prompt from step 1, filled in with:

    Text: “In the 1930s, drought struck the southern Great Plains. Farmers had plowed up the native grasses whose deep roots held the soil in place. When the rains failed, winds lifted the dry topsoil into dust storms. Drought alone did not cause the Dust Bowl; farming had left the land exposed.” Difficulty: long, complex sentences. Goal: explain how two causes combined. Key terms: drought (long dry spell), topsoil (top layer of soil), erosion (soil carried off by wind or water).

  2. AIEasier text, same claim

    Adapted text

    In the 1930s, drought struck the southern Great Plains. Drought is a long dry spell.

    Farmers had plowed up the native grasses. The grasses had deep roots. The roots held the soil in place.

    Topsoil is the top layer of soil. When the rains failed, winds lifted the dry topsoil into dust storms. This loss of soil is called erosion: soil carried off by wind or water.

    Drought alone did not cause the Dust Bowl. Farming had also left the land exposed.

    Every change and why

    • Split the long sentence about the grasses into three short ones, because long sentences make the original hard to follow.
    • Split the last sentence at its semicolon and added “also,” so the second cause still connects.
    • Explained “drought” and “topsoil” next to the words, using your meanings.
    • “Erosion” is not in the original. I added it, with your meaning, at the end of the third paragraph.

    Question

    Why did the dust storms happen? Use both causes and explain how they worked together.

    Answer for you

    The drought dried out the soil, and plowing had removed the grass roots that held it in place, so the wind could lift the topsoil. The text says drought alone did not cause the Dust Bowl.

    What to noticeSplitting sentences can cut the links between causes. This version keeps them, but it doesn’t spell out how the two causes combined, because that is what the question asks students to work out.

    The instruction behind this result, underlined in step 1“Keep the main point, words that limit it such as ‘only’ or ‘alone,’ and words that connect causes and effects, such as ‘when’ or ‘so.’”

Why it works

Describing the reading difficulty tells the AI which wording needs to change; the learning goal tells it what thinking students must still do. Keeping words such as “alone” and “so” helps preserve the meaning. Compare the list of changes with the original to check that nothing important was lost or added.

Check the answer

  • The adapted text keeps the main point, limits such as “only,” and the links between causes and effects.
  • Any key term the original didn’t use is marked as added, with where it was added.
  • The question needs the thinking your goal names, such as explaining causes, not just finding a fact.

Follow up

After students use the adapted text:

Prompt 3 of 10

Build review quizzes spaced over weeks

Use it when you think“My class does well on the unit test, and a month later most of it is gone.”

Copy the prompt

You’ll need

  • Your notes for the unit
  • Grade, number of quizzes, minutes each

Student work and records are personal data, and students may own the copyright in their work. Use only AI tools your school, college or district allows, and leave out names and anything that could identify a student.

Replace each [bracket] with your own material

Write short review quizzes from this unit, spread over the weeks after it ends.

UNIT NOTES
[Your notes for the unit]

CLASS
[Grade, number of quizzes, minutes each]

Use only facts in my notes. Ask questions students must answer from memory, such as short answers, not true-or-false. Spread the quizzes over several weeks, such as one, three and six weeks after the unit ends. Bring each idea back in a later quiz, worded differently. Put all the answers in a separate answer key.

Give me: the schedule, the quizzes with no answers showing, then the answer key.

See it work

With this prompt

Example written for this page

  1. YouThe prompt from step 1, filled in with:

    Class: grade 7 science, three quizzes, 5 minutes each. Notes: Earth’s outer layer is broken into plates that move a few centimeters a year. Where plates move apart, at divergent boundaries, new crust forms, as at the Mid-Atlantic Ridge. Where plates push together, at convergent boundaries, mountains such as the Himalayas can form. Where plates slide past each other, at transform boundaries, earthquakes are common, as along the San Andreas Fault.

  2. AIThree quizzes, weeks apart

    Schedule

    • Quiz 1: one week after the unit ends.
    • Quiz 2: three weeks after it ends.
    • Quiz 3: six weeks after it ends.

    Quizzes 2 and 3 bring back ideas from earlier quizzes in new words, so every idea comes up at least twice.

    Quiz 1

    1. What forms where two plates move apart?
    2. What kind of boundary is the San Andreas Fault?
    3. About how fast do plates move?

    Quiz 2

    1. What can form where two plates push together?
    2. What kind of boundary is the Mid-Atlantic Ridge?
    3. Why are earthquakes common along the San Andreas Fault?

    Quiz 3

    1. How far does a plate move in one year?
    2. What kind of boundary formed the Himalayas?
    3. Name one place where plates move apart.

    Answer key

    • Quiz 1: (1) new crust, (2) transform, (3) a few centimeters a year.
    • Quiz 2: (1) mountains, such as the Himalayas, (2) divergent, (3) the plates slide past each other there.
    • Quiz 3: (1) a few centimeters, (2) convergent, (3) the Mid-Atlantic Ridge.

    What to noticeEach idea returns after a gap, and students have to pull it from memory again. Rewording the question means they recall the idea, not just the wording of last time’s question.

    The instruction behind this result, underlined in step 1“Bring each idea back in a later quiz, worded differently.”

Why it works

Short-answer questions from your notes make students recall what you taught, not just recognize it, which strengthens memory more than rereading. Spreading the quizzes over weeks, with earlier ideas coming back in new words, adds the spacing that slows forgetting. The separate key lets you print the quizzes as they are.

Check the answer

  • Every question can be answered from your notes alone.
  • Later quizzes bring back earlier ideas in new words, weeks apart.
  • No answers appear on the quiz sheets; they’re all in the separate key.

Follow up

After each quiz, with the results:

Prompt 4 of 10

Let students solve more steps on their own

Use it when you think“Students can follow an example with every step shown, but freeze on the next problem.”

Copy the prompt

You’ll need

  • Your method and an example you’ve checked
  • Grade and what students already know

Student work and records are personal data, and students may own the copyright in their work. Use only AI tools your school, college or district allows, and leave out names and anything that could identify a student.

Replace each [bracket] with your own material

Build a short practice set that moves my students from an example with every step shown to solving on their own.

METHOD AND EXAMPLE
[Your method and an example you’ve checked]

STUDENTS
[Grade and what students already know]

Use my example as problem 1, showing every step. In each new problem, leave one more step for students to do, until the last problem has no help. Change the numbers and the setting each time. Make the last problem look different from the others, so copying the pattern isn’t enough. Don’t add a step I haven’t taught. Put the answers in teacher notes, with a possible mistake behind a wrong answer and how to check it.

Give me: the problems as students will see them, then the teacher notes.

See it work

With this prompt

Example written for this page

  1. YouThe prompt from step 1, filled in with:

    Grade 7 math. Method for percent change: find the difference between the two amounts; divide it by the original amount; multiply by 100 and say whether it’s an increase or a decrease. Checked example: a school garden grew 20 kg of tomatoes last year and 26 kg this year. 26 − 20 = 6; 6 ÷ 20 = 0.3; 0.3 × 100 = 30%, an increase. Students can divide and multiply decimals.

  2. AISupport removed step by step

    Problem 1, every step shown

    A school garden grew 20 kg of tomatoes last year and 26 kg this year.
    Difference: 26 − 20 = 6
    Divide by the original: 6 ÷ 20 = 0.3
    Multiply by 100: 30%, an increase

    Problem 2, finish the last step

    A school club had 80 members in January and 92 in February.
    Difference: 92 − 80 = 12
    Divide by the original: 12 ÷ 80 = 0.15
    Multiply by 100, then state increase or decrease: ____

    Problem 3, finish the last two steps

    A runner’s weekly distance went from 25 km to 30 km.
    Difference: 30 − 25 = 5
    Divide by the original: ____
    Multiply by 100, then state increase or decrease: ____

    Problem 4, on your own

    Visitors to a museum fell from 1,000 in May to 600 in June. Find the percent change.

    Teacher notes

    • Answers: problem 2 is 15%, an increase. Problem 3 is 5 ÷ 25 = 0.2, so 20%, an increase. Problem 4 is 400 ÷ 1,000 = 0.4, so 40%, a decrease.
    • Problem 4 goes down, so the original amount is the larger one. An answer of about 67% may mean the student divided by the new amount: 400 ÷ 600. Ask the student to show their steps before deciding that this is the mistake.
    • Every step comes from your method.

    What to noticeIn problems 1 to 3 the original amount is the smaller one, so a student copying that pattern may divide by 600 in problem 4. That is one possible explanation to check against the student’s steps.

    The instruction behind this result, underlined in step 1“Put the answers in teacher notes, with a possible mistake behind a wrong answer and how to check it.”

Why it works

An example you have checked gives the AI a correct starting point, but you still need to check the new problems and answers. Each new problem leaves another step for students to complete, until they solve one alone. A different-looking final problem helps you see whether students can use the method beyond the earlier pattern.

Check the answer

  • Each problem leaves one more step to the student than the one before.
  • The last problem looks different but needs only steps from your method.
  • Every answer in the teacher notes is right when you work it yourself.

Follow up

After students try the set:

Prompt 5 of 10

Build practice for telling concepts apart

Use it when you think“They can recite both rules and still pick the wrong one every time.”

Copy the prompt

You’ll need

  • Two concepts and the rule for telling them apart
  • Grade and the examples you taught

Student work and records are personal data, and students may own the copyright in their work. Use only AI tools your school, college or district allows, and leave out names and anything that could identify a student.

Replace each [bracket] with your own material

Write practice that helps my students tell two concepts apart.

CONCEPTS
[Two concepts and the rule for telling them apart]

STUDENTS
[Grade and the examples you taught]

First, write two pairs of questions that are the same except for the feature that decides the answer. Then write six mixed questions in an unpredictable order, with no headings or hints. Don’t let position or wording give the answer away. Use new examples, not the ones I taught. Ask students to give their answer and the words that decided it. If my rule would give the wrong answer for a question, leave it out and tell me.

Give me: the questions as students will see them, then an answer key, then any problem with my rule.

See it work

With this prompt

Example written for this page

  1. YouThe prompt from step 1, filled in with:

    Grade 4 grammar. Rule: use I when the pronoun does the action; use me when the action is done to it, or after words like to, for, with and between. Students can recite this but still write “Me and my friend went” and “for my friend and I.” Taught examples: “I ran home.” “Mom called me.”

  2. AIPairs, then mixed practice

    Practice

    Write I or me in each blank. Then underline the words that told you.

    1. My sister and ___ thanked the coach.
    2. The coach thanked my sister and ___.
    3. My cousin and ___ baked bread for Grandpa.
    4. Grandpa baked bread for my cousin and ___.
    5. Between you and ___, the movie was too long.
    6. After lunch, my friends and ___ cleaned the tables.
    7. ___ and my dad fixed the flat tire.
    8. Can you sit next to my brother and ___?
    9. The dog followed my neighbor and ___ home.
    10. Later, the twins and ___ walked home together.

    Answer key

    • I: 1 (thanked), 3 (baked), 6 (cleaned), 7 (fixed), 10 (walked). The pronoun does the action.
    • me: 2 (thanked) and 9 (followed), where the action is done to it; 4 (after for), 5 (after between), 8 (after next to).
    • Questions 1 to 4 are the two pairs. Question 7 puts the pronoun first, as in “Me and my friend went.”

    A problem with your rule

    Your rule depends on who does the action, so it doesn’t fit sentences like “My friend and ___ were picked” or “My friend and ___ are hungry,” which both need I. I left sentences like these out. You could add that the pronoun takes I whenever it’s the subject.

    What to noticeQuestions 1 and 2 use the same words with the roles swapped, so only the deciding feature changes the answer. In the mixed questions the blank moves around the sentence, so position can’t give the answer away.

    The instruction behind this result, underlined in step 1“First, write two pairs of questions that are the same except for the feature that decides the answer.”

Why it works

Pairs that differ only in the deciding feature make that feature the one thing that changes the answer. A mixed set with no pattern makes students apply the rule every time. Asking for the deciding words shows whether they used the rule or guessed, and checking each question catches sentences your rule can’t handle.

Check the answer

  • Each pair uses the same words, with only the deciding feature changed.
  • The mixed questions have no headings, and neither their order nor the blank’s position gives the answer away.
  • Every answer in the key follows your rule, and any sentence your rule can’t handle is flagged.

Follow up

After students try the practice:

Prompt 6 of 10

Turn notes into feedback students can use

Use it when you think“My comments keep saying ‘add more detail,’ and students don’t know what to change.”

Copy the prompt

You’ll need

  • The task, grading requirements and any source
  • The student’s work, without a name
  • What you noticed

Student work and records are personal data, and students may own the copyright in their work. Use only AI tools your school, college or district allows, and leave out names and anything that could identify a student.

Replace each [bracket] with your own material

Help me turn my notes on a student’s work into feedback they can act on.

TASK AND SOURCE
[The task, grading requirements and any source]

STUDENT WORK
[The student’s work, without a name]

MY NOTES
[What you noticed]

Use only the judgments in my notes. Open with something my notes say went well. For the one gap most worth fixing first, quote the student’s words and name the grading requirement. Give one revision step that says where to look and what kind of change to make. Don’t write the new sentence or choose the evidence for them. Tell me separately about any problem my notes miss, or any judgment you worded differently.

Give me: the feedback as the student will read it, then a note for me.

See it work

With this prompt

Example written for this page

  1. YouThe prompt from step 1, filled in with:

    Grade 7 English. Task: argue whether our school should switch to a four-day week, using the article. Criteria: a clear claim; evidence from the article; an explanation of how the evidence supports the claim. Article: in one district’s trial, attendance rose from 91% to 94%, and test scores stayed about the same. Work: “Our school should switch to a four-day week. Lots of students like it better. Four-day weeks are good.” My notes: claim clear; evidence vague; no explanation.

  2. AIOne change the student can try

    Draft feedback for the student

    You state your claim clearly in your first sentence: our school should switch to a four-day week.

    Next step: your second sentence says “Lots of students like it better.” The criteria ask for evidence from the article, and the article doesn’t say what students like. Go back to the article, choose the detail that best supports your claim, and use it in place of that sentence.

    For you

    • Your notes say the evidence is vague; I described it as not from the article, since the article doesn’t mention what students like. Change this if you meant something else.
    • I chose the evidence gap first, because the explanation will depend on which detail the student picks. Explaining how it supports the claim can be the next step.
    • Not in your notes: “Four-day weeks are good” repeats the claim without adding to it. I left it out of the feedback.

    What to noticeAsked plainly, the AI quoted both facts for the student. This step sends them back to the article to find and choose the detail themselves.

    The instruction behind this result, underlined in step 1“Don’t write the new sentence or choose the evidence for them.”

Without this prompt

A real AI reply, shortened

Replies “without this prompt” are shown only in the nine lessons where they fell short.

  1. YouAsked plainly, with the same material

    “Write feedback on this student’s work.”

  2. AIReplied

    Evidence from the article: not there yet. … The article does give you two facts you can use: in one district’s trial, attendance rose from 91% to 94%, and test scores stayed about the same. … Your next step: rewrite your second and third sentences…

Why it works

Starting from your notes keeps every judgment yours, and quoting the student’s words ties the feedback to something you can check against the work. Asked plainly, the AI handed the student the evidence; a step that says where to look leaves finding and choosing it to the student. Anything the AI notices beyond your notes, or words differently, is reported to you instead.

Check the answer

  • Every judgment in the feedback comes from your notes, and any rewording is pointed out to you.
  • The quoted words appear exactly in the student’s work.
  • The step says where to look and what to change, without writing the sentence or picking the evidence.

Follow up

When the student hands in the revision:

Prompt 7 of 10

Find out why students got it wrong

Use it when you think“Several students got it wrong, and I can’t tell whether it’s one mistake or several.”

Copy the prompt

You’ll need

  • The question and the method you taught
  • Students’ answers, with any work shown

Student work and records are personal data, and students may own the copyright in their work. Use only AI tools your school, college or district allows, and leave out names and anything that could identify a student.

Replace each [bracket] with your own material

Help me work out what these wrong answers show before I teach the topic again.

QUESTION AND METHOD
[The question and the method you taught]

ANSWERS
[Students’ answers, with any work shown]

Group the answers. For each wrong answer, suggest a possible mistake and show how it could produce that answer. Add a second possibility only if it also produces that answer. Say when you cannot suggest an explanation. These are possibilities to check, not confirmed reasons. Ask students one question about their steps without teaching the method. Then write one new problem at least as hard as the original, with its answer for me.

Give me: a table of the answers, then the question, then the new problem.

See it work

With this prompt

Example written for this page

  1. YouThe prompt from step 1, filled in with:

    Grade 5 end-of-lesson question: 1/2 + 1/3. This week I taught rewriting fractions with the same denominator (bottom number), using fraction bars. Students didn’t show their work. Answers from 24 students: 5/6 (11 students), 2/5 (7), 5/12 (3), 2/6 (2), 3/5 (1).

  2. AIPossible mistakes to check

    AnswerStudentsPossible explanationHow that mistake produces the answer
    5/611Correct answer; their steps were not shown3/6 + 2/6 = 5/6 is the taught method
    2/57Added the top numbers and the bottom numbers separately(1 + 1)/(2 + 3) = 2/5
    5/123Rewrote the fractions as 3/6 and 2/6, then added the bottom numbers too(3 + 2)/(6 + 6) = 5/12
    2/62Changed the bottom numbers to 6 but kept the top numbers; or added the original top numbers and multiplied the bottom numbers1/6 + 1/6 = 2/6; also (1 + 1)/(2 × 3) = 2/6
    3/51No explanation suggested from the answer aloneAsk this student to show the steps

    These calculations fit the answers, but they do not prove which steps the students took. If the suggested explanations are right, the 2/5 and 5/12 groups both added bottom numbers when adding fractions.

    Question to ask

    To each student: “How did you get the top and bottom numbers of your answer? Write a sentence for each.”

    New problem

    1/3 + 1/4. Like the original, it needs both fractions rewritten with a common denominator.

    Answer for you: 4/12 + 3/12 = 7/12. The same possible mistakes would produce 2/7 from adding the original tops and bottoms, 7/24 from adding the rewritten denominators, and 2/12 from changing the bottom numbers but not the top numbers.

    What to noticeAdding the original top and bottom numbers produces 2/5, but the answer alone does not show that this is what a student did. The follow-up question asks for the missing steps.

    The instruction behind this result, underlined in step 1“For each wrong answer, suggest a possible mistake and show how it could produce that answer.”

Why it works

A wrong answer can fit more than one mistake, so the AI shows possible calculations instead of claiming to know what the student did. Ask students to explain their steps before deciding what to teach again. A new problem at least as hard as the original then lets you check whether they can use the method.

Check the answer

  • Each suggested mistake produces the recorded answer, and is described as a possibility to check with the student.
  • The question asks how students got their answer, without teaching the method.
  • The new problem needs the same steps as the original, such as rewriting both fractions with a common denominator.

Follow up

When students answer your question:

Prompt 8 of 10

Write to a family about a concern

Use it when you think“I need to raise a concern with a family, and I don’t want it to sound like blame.”

Copy the prompt

You’ll need

  • What happened, with dates
  • The next step you’d like

Student work and records are personal data, and students may own the copyright in their work. Use only AI tools your school, college or district allows, and leave out names and anything that could identify a student.

Before pasting notes, remove reports of things you didn’t see and details about health or home life. Use your school’s messaging system to send the finished message to contacts on the student’s record.

Replace each [bracket] with your own material

Help me write a short message, under 150 words, to a student’s family about a concern, sticking to what happened.

MY NOTES
[What happened, with dates]

NEXT STEP
[The next step you’d like]

Use only facts from my notes, with their dates. Start with something going well, if my notes show one. Leave out labels, such as “lazy” or “rude,” and anything reported by someone else or about health or home life. Instead of my guesses about the cause, ask what the family has noticed. End with the next step. Write (student’s name) and (parent’s name) where names go.

Give me: the message, then a list of what you left out and why.

See it work

With this prompt

Example written for this page

  1. YouThe prompt from step 1, filled in with:

    Grade 4, Student A. Since Sept. 28, missed 4 of 6 math homework assignments. Head down on desk during math on Oct. 13 and Oct. 15. Quiz scores: 9/10 on Sept. 18, 5/10 on Oct. 9, 6/10 on Oct. 16. Seems lazy lately, probably lost interest in math. Next step: a 15-minute phone call this week.

  2. AIFacts, then an open question

    Draft message to the family

    Dear (parent’s name),

    (Student’s name) did well on the September 18 math quiz, scoring 9/10. Since September 28, though, they have missed 4 of the 6 math homework assignments. During math on October 13 and 15, they put their head down on the desk. Their quiz scores were 5/10 on October 9 and 6/10 on October 16.

    What have you noticed at home about how math and homework are going? You know your child best, and I’d like us to work on this together.

    Could we have a 15-minute phone call this week? Please let me know a time that suits you.

    Best wishes,
    (your name)

    Left out, and why

    • “Seems lazy”: a label, not something you saw.
    • “Probably lost interest in math”: a guess about the cause. The message asks what the family has noticed instead.

    What to noticeYour guess about the cause is left out, and an open question invites the family’s view instead.

    The instruction behind this result, underlined in step 1“Instead of my guesses about the cause, ask what the family has noticed.”

Why it works

Dated facts give the family something specific to respond to. The instructions make the rule against labels explicit, but you still need to check the whole draft against your notes. The open question asks what the family knows, and the list shows what was left out.

Check the answer

  • Every fact in the message, including each date and score, matches your notes.
  • No label or guess appears in the message, and the list names each one left out.
  • The message asks what the family has noticed and ends with one clear next step.

Follow up

When the family replies:

Prompt 9 of 10

Check a test before students take it

Use it when you think“I don’t want to find out a question was ambiguous after they’ve taken the test.”

Copy the prompt

You’ll need

  • The quiz or test
  • What students were taught

Student work and records are personal data, and students may own the copyright in their work. Use only AI tools your school, college or district allows, and leave out names and anything that could identify a student.

Don’t paste secure state tests, licensed test-bank questions or exams your school keeps confidential.

Replace each [bracket] with your own material

Check this quiz before my students take it. I’ll send my answer key, the answers I expect, after you’ve answered.

QUIZ
[The quiz or test]

TAUGHT MATERIAL
[What students were taught]

First, answer every question using only the taught material. Say so if the taught material supports more than one answer, or if a question needs something that wasn’t taught. Then stop and wait for my key.

When I send it, list every disagreement first. Then check each question for clues that give the answer away and for wording harder than the subject needs. Suggest the smallest fix for each problem, without making the question easier.

Give me: your answers first; after my key, a table of problems and fixes, most serious first.

See it work

With this prompt

Example written for this page

  1. YouThe prompt from step 1, filled in with:

    Quiz:
    1. An ___ branch carries out laws. (a) judicial (b) legislative (c) executive
    2. Which branch can stop a law? (a) executive (b) judicial (c) legislative
    3. To override a veto, each house needs: (a) a majority (b) two-thirds (c) three-quarters
    4. How long is a senator’s term?
    Taught: Congress (legislative) makes laws. The president (executive) carries them out and can veto bills. Two-thirds of both houses can override a veto. The Supreme Court (judicial) can strike down laws.

  2. AIAnswers before the key

    My answers, from the taught material only

    1. (c) executive.
    2. The taught material supports both (a) and (b): the president can veto bills, and the Supreme Court can strike down laws.
    3. (b) two-thirds. But (a) also fits the wording, since two-thirds of each house is also a majority.
    4. I can’t answer this. Senators’ terms aren’t in the taught material.

    Send me your key when you’re ready.

    What to noticeThe AI answered before seeing your key, so its answers are independent of yours. Its doubts about questions 2 and 3, and its flag on question 4, are already findings.

    The instruction behind this result, underlined in step 1“Say so if the taught material supports more than one answer, or if a question needs something that wasn’t taught.”

  3. YouSends the key

    My key: 1 (c), 2 (a), 3 (c), 4 six years.

    What to noticeThe key arrives only after the AI has committed to its own answers.

  4. AIDisagreements first

    Disagreements

    • Question 2: your key says (a), but the taught material supports (a) and (b).
    • Question 3: your key says (c) three-quarters; the taught material says two-thirds, and (a) also fits the wording.
    • Question 4: your key gives six years, but term lengths aren’t in the taught material.

    Problems and fixes, most serious first

    QuestionProblemSmallest fix
    3The key contradicts what was taughtChange the key to (b)
    2The taught material supports two answersAsk “Which branch can veto a bill?”
    4Tests something that wasn’t taughtTeach term lengths first, or replace the question
    3“A majority” is also true, since two-thirds is a majorityAsk “What minimum fraction of each house must vote to override a veto?” Use (a) just over one-half, (b) two-thirds, (c) three-quarters
    1“An” fits only “executive,” which gives the answer awayWrite “the ___ branch”

    What to noticeSix years is true, so a check against general knowledge would pass it. Checking against the taught material shows that students had no chance to learn it in class.

Why it works

Answering before it sees your key gives you a second, independent set of answers. Any question where the two sets differ stands out, including ones you wouldn’t think to question. A model can lean toward an answer it’s shown, so holding the key back keeps the comparison honest.

Check the answer

  • The AI answered every question, then stopped and waited for your key.
  • Disagreements with your key come before any other comment.
  • Each fix changes as little as possible and doesn’t make the question easier.

Follow up

After you revise the quiz, open a new chat so the AI hasn’t seen your key, and send:

Prompt 10 of 10

Test your assignment against AI

Use it when you think“Could a student just get AI to do this assignment, even one built on our class material?”

Copy the prompt

You’ll need

  • The assignment instructions
  • How you will grade the assignment

Student work and records are personal data, and students may own the copyright in their work. Use only AI tools your school, college or district allows, and leave out names and anything that could identify a student.

Start each attempt in a new chat without saved memory or shared files containing the class material. Keep that material out of the first chat, and use the same model and settings for both.

Replace each [bracket] with your own material

Attempt this assignment using only the assignment instructions and grading requirements below. This is the first attempt; I have not supplied the class material.

ASSIGNMENT
[The assignment instructions]

GRADING REQUIREMENTS
[How you will grade the assignment]

Write the complete answer, keeping to the assignment’s word limit. Don’t invent facts or figures from missing class material. If that stops you finishing, give the parts you can write, then name what is missing outside the answer. Don’t grade your own work.

Give me: the full attempt, then any missing material.

See it work

With this prompt

Example written for this page

  1. YouThe prompt from step 1, filled in with:

    First chat, assignment instructions only. First-year environmental studies. Assignment: In no more than 100 words, use the campus cup audit to recommend whether the cafeteria should stop selling drinks in single-use cups. Grading requirements: A clear recommendation; two figures from the audit; one limitation of the audit.

  2. AIFirst chat: full attempt

    Attempt A: instructions only

    I recommend keeping single-use cups for now and reviewing the campus audit before deciding whether to remove them. The decision should weigh cup use, waste and whether diners could use reusable cups. Without the audit, I cannot give its figures or assess how representative it is. Keeping the current arrangement would be a temporary choice while the required evidence is checked.

    Missing material

    The campus cup audit, including its figures and how it was collected.

    What to noticeThis chat has not received the audit. The answer gives a recommendation but names the missing evidence instead of inventing figures.

    The instruction behind this result, underlined in step 1“Don’t invent facts or figures from missing class material.”

  3. YouSecond new chat: full request

    Attempt this assignment using the assignment instructions, grading requirements and class material below. Write the complete answer within the stated length. Don’t invent facts or figures. If anything is missing, say so after the answer. Don’t grade your work.

    ASSIGNMENT
    In no more than 100 words, use the campus cup audit to recommend whether the cafeteria should stop selling drinks in single-use cups.

    GRADING REQUIREMENTS
    A clear recommendation; two figures from the audit; one limitation of the audit.

    CLASS MATERIAL
    Campus cup audit (main cafeteria, one week in March): 4,600 single-use cups used, 9% recycled; in a survey of 150 diners, 62% said they would bring their own cup to save 25 cents.

    What to noticeThis is the first message in a separate new chat with the same model and settings, not a reply to attempt A. The teacher supplies the identical assignment instructions and grading requirements, now with the audit.

  4. AISecond chat: full attempt

    Attempt B: with the audit

    I recommend a trial of ending single-use cup sales, with a reusable option for diners who arrive without a cup. The cafeteria used 4,600 single-use cups in one week, and only 9% were recycled. That gives us a reason to reduce disposable cup use. However, the audit covers one cafeteria for one week, so it may not represent other sites or seasons. The survey’s 62% willingness to bring a cup is a stated intention, not proof of what diners will do.

    What to noticeThe full answer uses the supplied figures and identifies the audit’s limited scope. The teacher can now compare both complete attempts against the same grading requirements.

  5. YouMy assessment and next step

    My check of both saved answers

    Grading requirementAttempt A: instructions onlyAttempt B: with the audit
    Clear recommendationMet: “keeping single-use cups for now”Met: “a trial of ending single-use cup sales”
    Two audit figuresNot met: no figures suppliedMet: “4,600 single-use cups” and “9% were recycled”
    One audit limitationNot met: says it cannot assess the auditMet: “one cafeteria for one week”
    No more than 100 wordsMet: 61 words in the answerMet: 81 words in the answer

    Attempt B meets these grading requirements in this example. That does not make the assignment safe or unsafe for every AI tool.

    A change I would try

    Before writing, students spend five minutes in class choosing two audit figures and one limitation, with a sentence explaining each choice. I can then ask about their reasoning. I would shorten the written task to make room for that step and check that the workload still fits.

    What to noticeThe teacher, not the AI, makes these judgments from both full answers. The proposed classroom step makes reasoning observable without claiming that the assignment is AI-proof.

Why it works

Keeping the class material out of the first chat lets you compare an answer written without that material with an answer written using it. Save both complete answers and check them yourself against the same grading requirements, rather than relying on the AI’s scores. The comparison tests this AI model on this assignment and can guide a change, but it does not show what every student or AI could do.

Check the answer

  • Use two separate new chats with identical instructions and grading requirements; add class material only to the second.
  • Both complete answers are saved, and you check each grading requirement against their actual words, including length.
  • Any change asks students to show their own reasoning; two attempts do not establish an AI-proof assignment.

Follow up

After saving the first answer, open a new chat with the same model and settings, add the class material and send:

Customer Service

Prompt 1 of 10

Answer a customer using your policy

Use it when you think“I know what the policy says, but I’m not sure it covers this customer.”

Copy the prompt

You’ll need

  • The customer’s message and order details
  • The policy that applies

Leave out names, contact details and payment details the AI doesn’t need, and follow your organization’s rules on what you can share.

Replace each [bracket] with your own material

Help me reply to this customer using only the facts and policy below.

CUSTOMER
[The customer’s message and order details]

POLICY
[The policy that applies]

First, list every condition in the policy that applies to this request, including exceptions, and mark each one met, not met or unknown. Check anything the customer assumes against the policy too.

Then write a reply of 120 words or fewer. Say which requests the policy allows. For anything it doesn’t allow, offer an allowed alternative if there is one. If an unknown condition would change the answer and we can’t look it up, ask the customer for it. Promise only what the policy allows. If the item might be faulty, tell me so; don’t imply the store policy is the limit of their rights.

Give me: the conditions list, then the reply.

See it work

With this prompt

Example written for this page

  1. YouThe prompt from step 1, filled in with:

    Customer: “These boots are too small. Can I swap them for a size up? I assume shipping is free, like last time.” Order: boots bought 12 days ago from the clearance section. Policy: exchanges are accepted within 30 days, with free return shipping. Clearance items can be exchanged, but the customer pays $7.95 return shipping.

  2. AIConditions, then the reply

    Conditions

    • Exchange within 30 days: met, bought 12 days ago.
    • Clearance item, so the customer pays $7.95 return shipping: met, bought from the clearance section.

    Draft reply

    Happy to help you get the right size. You’re within our 30-day exchange window, so you can swap the boots for a size up. Because they were bought on clearance, return shipping isn’t free this time: it’s $7.95. If that works for you, tell me the size you’d like and I’ll send the exchange instructions.

    What to noticeThe customer assumed free shipping, and the general policy seems to agree. Checking each condition catches the clearance exception before the reply promises free shipping.

    The instruction behind this result, underlined in step 1“First, list every condition in the policy that applies to this request, including exceptions, and mark each one met, not met or unknown.”

Why it works

Listing each condition, including exceptions, shows you at a glance whether the AI applied the whole policy or just the rule the customer asked about. That matters most with long policies, where exceptions are easy to miss. Possibly faulty items are flagged because, in many countries, consumer law gives rights for faulty goods that a store policy can’t remove.

Check the answer

  • Every condition that applies, including exceptions, appears in the list.
  • Every promise in the reply, including costs and time limits, matches the policy.
  • If the item might be faulty, the reply doesn’t present your policy as the customer’s only option.

Follow up

If the customer replies with new information:

Prompt 2 of 10

Ask the one question that settles the case

Use it when you think“I don’t know what to ask, and I don’t want to send the customer a checklist.”

Copy the prompt

You’ll need

  • The customer’s message and what we know
  • Available options and when each applies

Leave out names, contact details and payment details the AI doesn’t need, and follow your organization’s rules on what you can share.

Replace each [bracket] with your own material

Help me choose what to ask this customer next.

CASE
[The customer’s message and what we know]

OPTIONS
[Available options and when each applies]

Name the fact that decides between the options that could apply. Treat a fact as something we can look up only if my notes say so. Ask the customer for anything else we need. Ask the one question whose answer changes what I do next. If you can tell the customer what each answer would mean for them, do it in the same message, so they reply only if they need to. If the facts already show which option applies, name the deciding fact and skip the question.

Give me: the deciding fact, the message to send, then what to do for each answer.

See it work

With this prompt

Example written for this page

  1. YouThe prompt from step 1, filled in with:

    Customer: “You charged me twice for my order!” Our system shows one completed payment of $64, and we can’t see customers’ bank statements. Options: if the second charge is pending, it’s a temporary hold the bank will release, so no refund is needed. If both charges have gone through, we refund the duplicate.

  2. AIOne fact, both outcomes

    Deciding fact

    Whether the second charge on the customer’s statement is pending or has gone through. Your notes show one completed payment and say you can’t see bank statements, so only the customer can check this.

    Draft message to the customer

    Thanks for letting us know. On our side, I can see one completed payment of $64. If the second charge on your bank statement is marked pending, it’s a temporary hold your bank will release, and you don’t need to do anything. If it has gone through, reply here and we’ll refund it.

    What to do next

    • No reply, or they confirm it’s pending: nothing more to do.
    • They say it has gone through: find the second payment in your payment system and refund it.
    • They’re not sure: ask how the charge appears on the statement, with its date.

    What to noticeOne fact decides the case, and only the customer can see it. Explaining what each answer means lets a customer with a pending hold stop there, without writing back.

    The instruction behind this result, underlined in step 1“If you can tell the customer what each answer would mean for them, do it in the same message, so they reply only if they need to.”

Without this prompt

A real AI reply, shortened

Replies “without this prompt” are shown only in the nine lessons where they fell short.

  1. YouAsked plainly, with the same material

    “What should I ask this customer?”

  2. AIReplied

    The one thing you need to find out is whether the second charge is still pending or has actually posted. … Here’s a message you could send: … 2. What are the amount and date of each charge? 3. … send a screenshot…

Why it works

Asked plainly, an AI may draft a list of questions, and customers often answer long checklists slowly or not at all. Asking for the one question whose answer changes what you do keeps the message short. If the message explains what each answer means, the customer replies only when they need to.

Check the answer

  • Different answers to your question lead to different actions.
  • The message doesn’t ask for anything you could look up yourself.
  • Every refund or other promise comes from your options, not from the AI.

Follow up

If the customer writes back:

Prompt 3 of 10

Calm an upset customer without overpromising

Use it when you think“The customer is furious, and I can’t give them what they’re asking for.”

Copy the prompt

You’ll need

  • The customer’s message and what happened
  • What you’re allowed to offer

Leave out names, contact details and payment details the AI doesn’t need, and follow your organization’s rules on what you can share.

Replace each [bracket] with your own material

Help me reply to this complaint. I want to calm things down without promising anything I can’t deliver.

COMPLAINT
[The customer’s message and what happened]

MY OPTIONS
[What you’re allowed to offer]

Apologize once, for the specific thing that went wrong, then move to what I can do. Offer only what is on my list, and say what each option gives them. If they asked for something that isn’t on the list, say so kindly and plainly. If they ask for a manager, explain how to reach one using only my list, and don’t promise the outcome. Don’t describe anything as done or approved unless it is.

Keep the reply under 130 words and end with one clear choice or next step.

See it work

With this prompt

Example written for this page

  1. YouThe prompt from step 1, filled in with:

    Customer: “That’s the second day off I’ve wasted. Send a new one today and give me $100 for my time, or I’m done with you.” Two deliveries were missed. I’m refunding the $18 delivery fee. I can also rebook for Thursday or Saturday, or hold the parcel at our Elm Street depot. Any credit needs a manager’s approval, which I’ll request today. I can’t arrange a replacement or same-day delivery.

  2. AIDraft reply to the customer

    I’m really sorry. You took two days off for deliveries that never came, and that shouldn’t have happened.

    I’m refunding your $18 delivery fee now, and I can get your parcel to you in one of three ways: delivery on Thursday, delivery on Saturday, or pickup from our Elm Street depot.

    I’m not able to arrange a replacement or same-day delivery. I can ask my manager to approve a $100 credit, but I can’t promise the outcome. I’ll send that request today.

    Which delivery option would you like?

    What to noticeYou can request a $100 credit, but only your manager can approve it. The draft explains this limit and offers the delivery options you can provide.

    The instruction behind this result, underlined in step 1“Don’t describe anything as done or approved unless it is.”

Why it works

Your list of options keeps the reply specific and honest: every offer is one you can deliver, and what you can’t do is said plainly. Research on handling complaints suggests customers care more about seeing their problem solved than about a long apology. So the prompt asks for one apology, then options.

Check the answer

  • Every offer in the reply is on your list.
  • The apology names what went wrong, once, and the reply then moves to options.
  • Nothing is described as done or approved before it is.

Follow up

If the customer pushes back:

Prompt 4 of 10

Check a draft reply before you send it

Use it when you think“The draft sounds fine, but I’m not sure every sentence in it is true.”

Copy the prompt

You’ll need

  • Your draft reply
  • The customer’s question and the facts
  • The policy that applies

Leave out names, contact details and payment details the AI doesn’t need, and follow your organization’s rules on what you can share.

Replace each [bracket] with your own material

Check this draft reply against the facts and the policy before I send it. I want a fact check, not a style edit.

DRAFT
[Your draft reply]

QUESTION AND FACTS
[The customer’s question and the facts]

POLICY
[The policy that applies]

Go through the draft one claim or promise at a time. For each, say whether the facts or the policy support it, show it is wrong, or leave it unconfirmed. Check that the draft answers what the customer actually asked.

Then give me a corrected reply that changes only what needs to change. If the right answer depends on something we don’t know, list what I should confirm before sending instead of guessing.

Give me: the check as a table, then the corrected reply.

See it work

With this prompt

Example written for this page

  1. YouThe prompt from step 1, filled in with:

    Draft: “Good news, your refund is approved and will be in your account tomorrow.” Question: “Has my refund been approved yet?” Facts: we received the return yesterday; it hasn’t been inspected. Policy: refunds are approved after inspection, which takes up to 2 business days. Approved refunds are sent within 3 business days, and banks can take longer to show them.

  2. AIFact check and correction

    Check

    Claim in the draftWhat the facts and policy sayResult
    “Your refund is approved”The return hasn’t been inspected, and approval comes after inspectionThe facts show this is wrong
    “In your account tomorrow”Refunds are sent within 3 business days of approval, and banks can take longerUnconfirmed

    The draft answers the customer’s question, but wrongly.

    Corrected draft reply

    Thanks for checking in. We received your return yesterday. We haven’t inspected it, so your refund isn’t approved yet. Inspection takes up to 2 business days. Once your refund is approved, we send it within 3 business days, though your bank may take longer to show it.

    What to noticeThe draft sounded helpful but said the opposite of the facts. Checking it claim by claim catches the error, and the correction ties each time limit to the step it belongs to.

    The instruction behind this result, underlined in step 1“For each, say whether the facts or the policy support it, show it is wrong, or leave it unconfirmed.”

Why it works

Asked to “improve” a draft, an AI may polish the wording without checking the facts. Asking for a claim-by-claim check against your facts makes it test each statement, and the table shows you exactly what it checked.

Check the answer

  • Each claim in the table is matched to a specific fact or policy line.
  • The corrected reply answers the question the customer actually asked.
  • Each date or time limit is tied to the step it really belongs to.

Follow up

After you edit the reply yourself:

Prompt 5 of 10

Plan troubleshooting steps from your guide

Use it when you think“The customer has tried the obvious fixes, and I’m not sure what to check next.”

Copy the prompt

You’ll need

  • The problem and what’s been tried
  • Your troubleshooting guide

Leave out names, contact details and payment details the AI doesn’t need, and follow your organization’s rules on what you can share.

Replace each [bracket] with your own material

Help me plan the next troubleshooting steps for this customer, using only our guide.

PROBLEM
[The problem and what’s been tried]

GUIDE
[Your troubleshooting guide]

Skip anything already tried. Start with the check that best narrows down the cause. For each possible result, say what to do next and when to stop and pass the case to the next support team. Use only steps from the guide; if it gives no next step, say what information to collect for the next team. If the guide warns against something the customer might do, such as retrying, include a line telling them not to.

Give me: the first thing to ask the customer, then a short table of results and next steps.

See it work

With this prompt

Example written for this page

  1. YouThe prompt from step 1, filled in with:

    Customer’s spreadsheet file (CSV): “My CSV upload keeps failing, and there’s no error message.” Browser already restarted. Guide: uploads fail above 25 MB (megabytes) or without Import permission. Over 25 MB: split the file. 25 MB or under: an administrator (who manages account access) checks Import permission. If both checks pass, send the case, file size and upload time to the technical team. Repeated retries can create duplicate records.

  2. AINext steps from the guide

    First, ask the customer

    How big is the file?

    IfDo thisStop here?
    Over 25 MBSplit it into files of 25 MB or under; confirm each size, then upload each one onceStop if uploads work. If a confirmed file of 25 MB or under still fails, go to the next row
    25 MB or underAsk their account administrator to check Import permissionPause until permission is confirmed. If it is missing or unknown, the guide gives no next step: ask the account administrator for an approved next step.
    25 MB or under, permission confirmed, and upload still failsSend the case, file size and upload time to the technical teamYes, pass to the technical team

    Draft message to the customer: While we sort this out, please don’t keep retrying the same upload. Repeat attempts can create duplicate records.

    What to noticeThe plan pauses when permission is missing or unknown, because the guide gives no fix for that result. It also warns against repeat uploads, which can create duplicates.

    The instruction behind this result, underlined in step 1“If the guide warns against something the customer might do, such as retrying, include a line telling them not to.”

Why it works

The prompt asks the AI to use your guide and skip fixes the customer has already tried. This keeps its suggestions tied to your product’s instructions. Asking what to do for each possible result shows when to pause or involve the next support team.

Check the answer

  • No step repeats something the customer has already tried.
  • Every step comes from your guide.
  • The plan pauses if the guide has no next step and transfers the case only when the required checks pass.

Follow up

When the customer tells you what they found:

Prompt 6 of 10

Answer the next question before it’s asked

Use it when you think“Customers often write back a few days after I mark their problem as solved.”

Copy the prompt

You’ll need

  • What was done for the customer
  • Billing or settings affected by the change

Leave out names, contact details and payment details the AI doesn’t need, and follow your organization’s rules on what you can share.

Replace each [bracket] with your own material

Write a short closing message for this customer that also answers the question they are most likely to ask next.

WHAT I DID
[What was done for the customer]

ACCOUNT DETAILS
[Billing or settings affected by the change]

First, name the most likely follow-up question and the detail above that makes it likely, such as a changed bill, a new setting or a step they still have to take. If nothing points to a likely question, say so and don’t add one.

Then write the message in 100 words or fewer: confirm what’s done, then answer that question with the exact details. If the details mention a page where they can check for themselves, include it.

Give me: the likely question and why, then the message.

See it work

With this prompt

Example written for this page

  1. YouThe prompt from step 1, filled in with:

    I upgraded the customer’s plan to Pro today. Their next bill, on November 1, will be $59.20: $45 for Pro plus $14.20 for the rest of October. After that, it’s $45 a month. They can see the breakdown under Billing → Invoices.

  2. AILikely question, then the reply

    Likely question

    “Why is my first bill higher than $45?” The November 1 bill includes an extra $14.20 for the rest of October.

    Draft reply

    Your Pro plan is now active. Your next bill, on November 1, will be $59.20: $45 for Pro plus a one-time $14.20 for the rest of October. After that, it’s $45 a month. You can see the breakdown any time under Billing → Invoices.

    What to noticeThe first bill is higher than the later monthly bills because it includes part of October. The draft explains the extra charge before the customer needs to ask.

    The instruction behind this result, underlined in step 1“First, name the most likely follow-up question and the detail above that makes it likely, such as a changed bill, a new setting or a step they still have to take.”

Why it works

Asking the AI to name the likely next question, and the detail behind it, makes your closing message answer that question up front. Every prediction has to point to a detail you supplied, so the message sticks to your facts.

Check the answer

  • The predicted question is tied to a specific detail, not a hunch.
  • The reply confirms what was done first, then answers the question with exact figures or steps.
  • Any page the reply points to is one you supplied.

Follow up

If the customer contacts you again after you close the case:

Prompt 7 of 10

Pass a case to a colleague with clear notes

Use it when you think“I’m passing this case to a colleague, and I don’t want the customer to have to explain it all again.”

Copy the prompt

You’ll need

  • The conversation and case notes
  • The colleague or team taking over

Leave out names, contact details and payment details the AI doesn’t need, and follow your organization’s rules on what you can share.

Replace each [bracket] with your own material

Write a note for the colleague taking over this customer’s case.

CASE
[The conversation and case notes]

HANDING TO
[The colleague or team taking over]

Start with two sentences: what the customer needs and what has to happen next. Then list what has happened so far, in time order; what was tried and what the result was; any promise made to the customer, with its date, time and who will keep it; and open questions. Keep exact error messages, IDs and times. Separate what someone did from what was confirmed to work. If the notes disagree, say so instead of picking one. Leave out empty sections.

Use short headings and bullet points.

See it work

With this prompt

Example written for this page

  1. YouThe prompt from step 1, filled in with:

    Monday, October 19, times in Eastern Time. 9:05 Customer reports that the PDF of invoice INV-20417 is blank. 9:12 I asked them to refresh; they say it’s still blank. 9:20 Billing notes say “invoice regenerated.” 9:25 Customer says it’s still blank. 9:30 I promised an update by 2:00 p.m. Handing to: the Billing team, who will send that update.

  2. AIDraft note to the next colleague

    Summary

    The customer needs a readable PDF of invoice INV-20417; the customer says it is still blank after refreshing and after Billing created a new file. Billing needs to find out why and send the customer an update by 2:00 p.m. ET today.

    What happened (Monday, October 19, ET)

    • 9:05 Customer reported that the PDF of invoice INV-20417 is blank.
    • 9:12 Asked them to refresh; they say it’s still blank.
    • 9:20 Billing’s notes say “invoice regenerated,” meaning a new invoice file was created.
    • 9:25 Customer says it’s still blank.
    • 9:30 Promised the customer an update by 2:00 p.m.

    Tried, and the result

    • Refresh: didn’t help.
    • New file created, but not confirmed to work: the customer still sees a blank file.

    Promised to the customer

    • An update by 2:00 p.m. ET on Monday, October 19. Billing will send it.

    Open question

    • Is the new file itself blank, or is the customer still opening the old one?

    What to notice“Invoice regenerated” records that someone created a new file; it doesn’t show that the problem was fixed. The note keeps the action separate from the result the customer reported.

    The instruction behind this result, underlined in step 1“Separate what someone did from what was confirmed to work.”

Why it works

In a long conversation, a summary can mistake “invoice regenerated,” meaning a new file was created, for “problem solved.” Separating what was done from what was confirmed to work shows the next colleague what still needs attention. Recording who will keep each promise and when makes those commitments easier to follow.

Check the answer

  • The first two sentences say what the customer needs and what happens next.
  • Every promise to the customer appears with its date, time and who will keep it.
  • Nothing is marked as fixed unless the notes show that it worked.

Follow up

Before you send the note, to catch anything missing:

Prompt 8 of 10

Turn a solved case into a help article

Use it when you think“I solved this problem, but the next support agent will have to work it out from scratch.”

Copy the prompt

You’ll need

  • Case notes, the fix and how it was checked
  • Product version, setup and permissions

Leave out names, contact details and payment details the AI doesn’t need, and follow your organization’s rules on what you can share.

Replace each [bracket] with your own material

Turn this solved case into a draft help article that support agents and customers can follow.

SOLVED CASE
[Case notes, the fix and how it was checked]

PRODUCT SETUP
[Product version, setup and permissions]

Write the title as the question a customer would search for. Describe the problem in general terms, not this case’s dates or numbers. Include who it applies to, the symptoms, numbered steps, and how to tell it worked. Claim the fix only for the version and setup in the case. Leave out steps that didn’t work, and remove the customer’s details. If a step is missing detail, such as which menus to open, mark it as a question for the reviewer instead of making it up.

End with a short reviewer checklist, including testing the steps and checking the whole article against our publishing requirements.

See it work

With this prompt

Example written for this page

  1. YouThe prompt from step 1, filled in with:

    In LedgerLite 4.2, a customer’s export was missing October invoices. Their saved date filter ended on September 30. Changing the end date to October 31 fixed it, and the export then matched the Invoices page. Exporting needs the Export permission. We haven’t tested other versions.

  2. AIDraft article

    Why are some invoices missing from my LedgerLite export?

    Applies to: LedgerLite 4.2 users with the Export permission who use a saved date filter.

    Symptoms: Recent invoices are missing from an export.

    Steps

    1. Open your saved date filter. (Reviewer: name the menus to open.)
    2. Check the end date. If it’s earlier than the invoices you need, change it to the last day you want included.
    3. Run the export again.

    How to tell it worked: The export should match the Invoices page for the same dates. (Reviewer: specify how to compare them; equal invoice counts alone don’t establish a match.)

    For the reviewer

    • Confirm which menus open the saved date filter.
    • Check whether other versions behave the same way.
    • Specify the comparison that checks the export against the Invoices page; the case doesn’t describe it.
    • Add where to get help if the export still doesn’t match.
    • Test all steps in version 4.2 and review the whole article against your publishing requirements.

    What to noticeThe case doesn’t say which menus to open or how the export was checked, so both are left for the reviewer. The article claims the fix only for version 4.2.

    The instruction behind this result, underlined in step 1“If a step is missing detail, such as which menus to open, mark it as a question for the reviewer instead of making it up.”

Why it works

Turning one case into a general article, an AI can fill missing steps with plausible menu names and imply the fix works everywhere. Telling it to claim only the tested version and to mark gaps for a reviewer keeps the article honest about what’s known. A reviewer should test the steps and check the whole article against your publishing requirements before it goes live.

Check the answer

  • The reviewer checklist includes testing the steps and reviewing the whole article, not just the flagged gaps.
  • The article claims the fix only for the version and setup in the case.
  • Every missing detail is marked for the reviewer, not invented.

Follow up

After the reviewer tests the steps and reviews the whole article:

Prompt 9 of 10

Decide which support request to answer first

Use it when you think“We share one inbox, and I end up answering whoever shouts loudest.”

Copy the prompt

You’ll need

  • Support requests, with arrival times
  • Your priority rules and response times
  • The current date and time

Leave out names, contact details and payment details the AI doesn’t need, and follow your organization’s rules on what you can share.

Replace each [bracket] with your own material

Put these support requests in the order I should answer them, using only my priority and response-time rules.

REQUESTS
[Support requests, with arrival times]

RULES
[Your priority rules and response times]

NOW
[The current date and time]

For each request, give its priority, the rule and request detail that determine it, the response deadline and the time left. Show the time arithmetic so I can check it. Ignore tone and customer size unless a rule mentions them. If a deadline depends on an assumption, such as which days count as business days, state it. If a request lacks what’s needed to set its priority, say what’s missing instead of guessing.

Give me: a table in the order I should work, then the arithmetic.

See it work

With this prompt

Example written for this page

  1. YouThe prompt from step 1, filled in with:

    It’s Monday, October 19, 10:00 a.m. Request A (arrived 9:50 a.m.): no one can place orders. Request B (8:00 a.m.): two receipts show the wrong tax. Request C (9:00 a.m.), from our largest customer: “URGENT!!! Our CEO needs dark mode by Friday.” Rules: outages are P1, respond within 15 minutes. Partial problems are P2, respond within 4 hours. Feature requests are P3, respond by the next business day.

  2. AIRequests in priority order

    OrderRequestPriority and whyRespond byTime left
    1AP1: no one can place orders (outage)10:05 a.m. today5 minutes
    2BP2: wrong tax on two receipts (partial problem)noon today2 hours
    3CP3: dark mode is a feature request, however urgent it soundsTuesday, October 20 (assumed)Not set: your rules give no time

    Arithmetic

    • A: 9:50 + 15 minutes = 10:05; 10:05 − 10:00 = 5 minutes left.
    • B: 8:00 + 4 hours = 12:00 (assumes clock hours, not business hours); 12:00 − 10:00 = 2 hours left.
    • C: assumes Monday to Friday are business days. Your rules don’t say what time of day the deadline falls.

    What to noticeRequest A arrived last but is due first, and C’s shouting doesn’t move it up, because your rules don’t mention tone or customer size. The assumptions behind the deadlines are stated, not hidden.

    The instruction behind this result, underlined in step 1“Show the time arithmetic so I can check it.”

Why it works

Giving the AI your rules and the current time, and asking for the arithmetic, makes the order quick to check. This also keeps tone and customer size out of the order, unless your rules mention them. If your support software already tracks response deadlines, keep using its timers; this prompt helps check the order in which to respond.

Check the answer

  • Each priority names a rule and a detail from the request.
  • You’ve worked out two deadlines yourself; AI tools can slip on time arithmetic.
  • Every assumption, such as business days, is stated rather than hidden.

Follow up

When support requests change or new ones arrive:

Prompt 10 of 10

Find the problems customers keep reporting

Use it when you think“My manager wants to know what customers complain about, and I have 200 support requests.”

Copy the prompt

You’ll need

  • Feedback or support requests, each with an ID
  • The question you need answered

Leave out names, contact details and payment details the AI doesn’t need, and follow your organization’s rules on what you can share.

Replace each [bracket] with your own material

Find the recurring problems in this customer feedback.

FEEDBACK
[Feedback or support requests, each with an ID]

QUESTION
[The question you need answered]

Group the feedback by the problem described; one message can belong to more than one group. For each group, give the count, the IDs and one short quote. Count repeat messages about the same incident once, and say which you merged. Keep what customers reported apart from guesses about causes, and suggest what to check. Separately, list any message that needs action now, such as money taken twice. A repeat contact doesn’t prove no reply was sent. With more than about 50 messages, first list each ID with its group. Don’t present these counts as how often a problem happens overall.

Give me: a table of groups, then what to look into first and why.

See it work

With this prompt

Example written for this page

  1. YouThe prompt from step 1, filled in with:

    Feedback from October 12–18. Question: which problem should we look into first? F1: “Payment timed out.” F2: “Checkout froze when I paid.” F3: “Payment timed out and I was charged twice.” F4: “Delivery was slow.” F5: same customer and order as F2, asking again.

  2. AIGroups and what to check

    ProblemCountIDsQuote
    Payment failing or freezing3F1, F2, F3“Payment timed out.”
    Charged twice1F3“I was charged twice.”
    Slow delivery1F4“Delivery was slow.”

    F5 repeats F2 (same customer, same order), so I counted it once. F3 is in two groups, so the counts add up to more than the four distinct messages.

    Needs action now: F3 reports being charged twice. Check the transaction and refund the duplicate if it’s confirmed. F5 is the F2 customer asking again. Check the case history to see whether they are waiting for a reply or for a resolution.

    Look into first: payment failures. They appear in three of the four messages. Check the payment logs for these orders to see whether payments failed or went through without a confirmation. Treat the double charge as unconfirmed until the transaction records show it.

    Then: slow delivery. One report. Compare F4’s promised and actual delivery dates.

    What to noticeA customer’s report isn’t proof of what happened or why. The summary flags a possible double charge and a repeat contact, without assuming the causes or whether a reply has been sent.

    The instruction behind this result, underlined in step 1“Keep what customers reported apart from guesses about causes, and suggest what to check.”

Why it works

Listing the message IDs and counts lets you trace each group back to the feedback. Separating customer reports from possible causes gives you things to investigate, rather than treating a report as proof. The IDs also let you check counts that the AI may get wrong.

Check the answer

  • Each count matches its list of IDs.
  • Merged duplicates are named, and messages in two groups are pointed out.
  • The summary states a cause or says no reply was sent only when the notes show it.

Follow up

When you have records that could explain a reported problem:

Sales

Prompt 1 of 10

Research a company before you contact it

Use it when you think“I found news about this company, and I’m about to assume it means they have a problem we can solve.”

Copy the prompt

You’ll need

  • What you found, with sources and dates
  • Your offer and evidence for its results
  • Who you’re writing to, and their role

Pricing, discount limits, deal histories and requests for proposal (RFPs) are often confidential. Use an AI tool your company approves, share only what you’re allowed to, and replace names with roles.

Before you send, add anything your company’s rules require in emails to potential customers, such as an opt-out line.

Replace each [bracket] with your own material

Help me write a first email to this person, using only my research and the results we have evidence for.

WHAT I FOUND
[What you found, with sources and dates]

WHAT WE OFFER
[Your offer and evidence for its results]

RECIPIENT
[Who you’re writing to, and their role]

First, list the facts, each with its source and date. Then list each need we might guess at, the fact that suggests it, and another explanation that would mean they don’t need us. If no fact connects to what we offer, say so and don’t write the email.

Otherwise, write an email of 100 words or fewer that uses only the facts and asks about the need instead of assuming it. Fit the question to the recipient’s role, claim only results backed by our evidence, and end with one easy next step.

Give me: the facts, the guesses, then the email.

See it work

With this prompt

Example written for this page

  1. YouThe prompt from step 1, filled in with:

    Job post, Fernhill Home Care, October 2: “Scheduling coordinator. Build weekly schedules for 140 caregivers in Excel and cover last-minute absences by phone.” Local news, September 18: Fernhill won a county contract; its CEO said, “Finding enough caregivers is our biggest challenge.” Our product: RosterLeaf builds schedules from staff availability and texts open shifts to available caregivers. Approved case study: Lakeside Care cut scheduling time from 12 to 4 hours a week. Writing to: the director of operations.

  2. AIFacts, guesses, then the email

    Facts

    • Fernhill is hiring a scheduling coordinator to build weekly schedules for 140 caregivers in Excel and cover last-minute absences by phone (job post, October 2).
    • Fernhill won a county contract, and its CEO said finding enough caregivers is its biggest challenge (local news, September 18).
    • RosterLeaf builds schedules from staff availability and texts open shifts to available caregivers. Lakeside Care cut scheduling time from 12 to 4 hours a week (our approved case study).

    Guesses

    • Building schedules in Excel takes up a lot of time. Suggested by: the job post. Another explanation: Excel works fine, and the new role is mostly about covering absences.
    • Last-minute absences are hard to fill. Suggested by: the job post, which says absences are covered by phone. Another explanation: absences are rare, and phone calls work at their size.
    • RosterLeaf could ease their caregiver shortage. Suggested by: the CEO’s comment. Another explanation: the shortage is about recruiting, and nothing we can show helps with that, so the email shouldn’t claim it.

    Draft email

    Subject: Schedules and open shifts at Fernhill

    Hello,

    I saw Fernhill’s posting for a scheduling coordinator to build weekly schedules for 140 caregivers in Excel and cover last-minute absences by phone.

    RosterLeaf builds schedules from staff availability and texts open shifts to available caregivers. At Lakeside Care, it cut scheduling time from 12 to 4 hours a week.

    How quickly do open shifts get filled today, and is that something you’d like to change?

    If it’s useful, I can send you the Lakeside case study.

    What to noticeThe CEO’s hiring worry might seem a good reason to contact the company, but nothing shows RosterLeaf helps with recruiting. Listing it as a guess with another explanation keeps it out of the email, which asks about filling open shifts instead.

    The instruction behind this result, underlined in step 1“Then list each need we might guess at, the fact that suggests it, and another explanation that would mean they don’t need us.”

Why it works

Asking for each fact with its source, and each guessed need with another explanation, makes the AI show what the email rests on. You can check every line before it goes out. Listing the CEO’s hiring worry as a guess keeps it out of the email, since you can’t show that RosterLeaf solves it.

Check the answer

  • Every fact in the email appears in the facts list with its source, and its date if it has one.
  • Each guess has another explanation, and the email asks about the need instead of stating it.
  • Results in the email are ones you can show, and they aren’t promised to this company.

Follow up

When they reply:

Prompt 2 of 10

Prepare questions for a first sales call

Use it when you think“I have a first sales call this week, and I don’t know which answers would change what I propose.”

Copy the prompt

You’ll need

  • What you know about the customer
  • What you plan to offer, and its limits
  • Who will be on the call

Pricing, discount limits, deal histories and requests for proposal (RFPs) are often confidential. Use an AI tool your company approves, share only what you’re allowed to, and replace names with roles.

Replace each [bracket] with your own material

Help me choose the three questions that matter most for this first call about the customer’s needs.

WHAT WE KNOW
[What you know about the customer]

OUR PLAN
[What you plan to offer, and its limits]

ON THE CALL
[Who will be on the call]

Find the things our plan depends on that we haven’t confirmed, and skip anything we already know. Write three questions that test them, most important first. For each, give two realistic answers that would lead to different next steps, and say what we would do after each. Word each question so it doesn’t assume a problem or agreement, and say who on the call is best placed to answer it. End with the answer that would mean we should change our plan or walk away.

Give me: the three questions with their answers and next steps, then the answer that would change our plan.

See it work

With this prompt

Example written for this page

  1. YouThe prompt from step 1, filled in with:

    Web form from an insurance office: “We’d like charging stations for electric cars in our 60-space parking lot. About a dozen of our staff drive electric cars.” First sales call Thursday with their facilities manager and sustainability lead. Our plan: ten Level 2 chargers (a common workplace type), installed, with a booking app. Our price assumes the building already has enough electrical capacity, and we don’t do electrical upgrades.

  2. AIQuestions and where they lead

    1. Has anyone checked how much spare electrical capacity the building has? (Ask the facilities manager.)
    • Yes, and there’s room → use the confirmed capacity to decide how many chargers can be installed.
    • No one has checked → find out who can check it before we quote, since our price assumes there’s enough capacity already.
    1. Besides the dozen staff with electric cars, will company vehicles need to charge here too? (Ask the sustainability lead.)
    • No, just staff → keep the plan of ten chargers.
    • Yes → check the vehicles’ battery sizes and how long they’re parked overnight before proposing Level 2 chargers.
    1. Do you own the parking lot, or would a landlord need to approve the work? (Ask the facilities manager.)
    • They own it → no landlord approval is needed.
    • They lease it → find out who can approve the work on the landlord’s behalf, and involve them before we quote.

    Answer that would change our plan

    “The building has no spare capacity.” We don’t do electrical upgrades, so our package wouldn’t fit unless someone else added capacity first.

    What to noticeEach question has two answers that lead to different actions, and none asks what the web form already says. Asking about capacity first means you find out on the first call, not after the quote, whether the building can take the chargers.

    The instruction behind this result, underlined in step 1“For each, give two realistic answers that would lead to different next steps, and say what we would do after each.”

Why it works

Asking for two possible answers to each question, each with its own next step, shows what every question could change. You can drop any whose answer wouldn’t change your plan. Telling the AI what you know and who will attend keeps out questions you could answer yourself and aims each at the right person.

Check the answer

  • Each question’s two answers lead to different next steps.
  • Nothing is asked that you already know, and no question assumes a problem or agreement.
  • At least one possible answer would make you change the plan or walk away.

Follow up

After the call, with their answers:

Prompt 3 of 10

Recap a sales call without overstating it

Use it when you think“I need to write up today’s call without putting words in the customer’s mouth.”

Copy the prompt

You’ll need

  • Your call notes, with who said what
  • Who will read the recap

Pricing, discount limits, deal histories and requests for proposal (RFPs) are often confidential. Use an AI tool your company approves, share only what you’re allowed to, and replace names with roles.

Replace each [bracket] with your own material

Turn my call notes into a recap for the customer and a short brief for my team.

NOTES
[Your call notes, with who said what]

RECAP GOES TO
[Who will read the recap]

Keep every statement as certain as it was said: an estimate stays an estimate, and a hope stays a hope. In the recap, include only what the customer would expect to see in writing: their goals, needs, current costs, timing and agreed actions. Name who will do each action, and give a date only if one was agreed. Put budget discussions, internal opinions and our own assessment in the team brief, and say who said each point. End the brief with up to three open questions, split into questions for the customer and things to check on our side.

Give me: the recap, written as I’d send it, then the brief.

See it work

With this prompt

Example written for this page

  1. YouThe prompt from step 1, filled in with:

    Call with the operations manager at an architecture firm, Thursday, October 15. Printing: “probably around $2,000 a month, I’d have to check.” Scanning to project folders is a must; color on every floor would be nice. “Hoping to decide by the end of November.” The partners haven’t seen a budget, and one thinks leasing is a waste of money. She sends last quarter’s invoices by Friday; I quote after that. Recap to her; she may forward it to the partners.

  2. AIRecap, then the brief

    Draft recap

    Thanks for your time today. Here’s what I took away. Please correct anything I’ve got wrong.

    • Scanning to your project folders is a must, and color printing on every floor would be nice to have.
    • Your printing costs are probably around $2,000 a month, though you’d need to check.
    • You’re hoping to decide by the end of November.

    Next steps:

    • You’ll send last quarter’s invoices by Friday.
    • I’ll send a quote once I have them.

    Brief for our team

    Every point below comes from the operations manager unless it’s marked as our view.

    • Needs: scanning to project folders is a must; color on every floor is a nice-to-have.
    • Cost today: “probably around $2,000 a month, I’d have to check.” Invoices due Friday.
    • Timing: hoping to decide by the end of November. Not a deadline.
    • Budget: the partners haven’t seen one.
    • Others involved: the partners. One thinks leasing is a waste of money. Who gives final approval is unconfirmed.
    • Our view: lower costs and dependable scanning look like the goals, but she didn’t say so directly.

    Open questions

    For the customer:

    • Who makes the final decision, and what will the partners want to see?
    • Where are your project folders kept?

    To check on our side:

    • Can our printers scan straight to wherever their project folders are kept?

    What to noticeHer hope stays a hope, and her cost figure stays an estimate. The budget and the skeptical partner appear only in the brief for your team, so nothing awkward reaches the partners if she forwards the recap.

    The instruction behind this result, underlined in step 1“Keep every statement as certain as it was said: an estimate stays an estimate, and a hope stays a hope.”

Why it works

Summaries can drop small words like “probably” and “hoping.” Asking for each statement to stay as certain as it was said keeps them in, so you don’t put words in the customer’s mouth. Saying who will read the recap, and asking for a separate brief, keeps budget talk and opinions out of an email that may be forwarded.

Check the answer

  • Estimates, hopes and nice-to-haves read as they were said, not as facts, deadlines or needs.
  • The recap leaves out budget talk, internal opinions and your own view of the deal.
  • Each action names who will do it, and includes a date only if one was agreed.

Follow up

Before you send the recap, to catch anything overstated:

Prompt 4 of 10

Find out who makes the buying decision

Use it when you think“Everyone on the call nodded, and I still don’t know who decides or who will push for us.”

Copy the prompt

You’ll need

  • Notes and emails, with who said what
  • What you know about their buying process

Pricing, discount limits, deal histories and requests for proposal (RFPs) are often confidential. Use an AI tool your company approves, share only what you’re allowed to, and replace names with roles.

Replace each [bracket] with your own material

Help me map who is involved in this buying decision, using only these records.

RECORDS
[Notes and emails, with who said what]

PROCESS
[What you know about their buying process]

Make a table with these columns: person, what they have said or done, only attended or received a copy of an email (yes or no), role (confirmed or unknown), and what to ask next. Roles are: uses it, evaluates it, recommends it, can block it, approves the budget, or handles purchasing; one person can have several. Don’t give anyone a role because of their title, seniority or friendliness. If someone only attended or received an email without replying, mark their role unknown and ask about it. For anyone we hope will push for us, say what they have actually agreed to do.

Give me: the table, then up to three questions about their process that would close the biggest gaps.

See it work

With this prompt

Example written for this page

  1. YouThe prompt from step 1, filled in with:

    Video call, October 6: warehouse manager, safety lead, operations director. Warehouse manager: “I’ll test two of your forklift cameras on the night shift.” The safety lead asked how incident footage is stored. The operations director left after ten minutes without speaking. Email from the warehouse manager, October 8, copying the director and a finance contact: “Our director will want to see the trial results before anything goes to finance.” Their rule: spending over $20,000 needs finance approval. Our quote: $26,000.

  2. AIThe map, from the records

    PersonWhat they have said or doneOnly attended or received email?RoleWhat to ask next
    Warehouse managerAgreed to test two cameras on the night shift. Wrote that the director will want the results before financeNoEvaluates it: confirmed. Recommends it: unknown; hasn’t offered toWhat must the trial show, and who will see the results?
    Safety leadAsked how incident footage is storedNoUnknownDoes the safety lead need to approve how footage is stored?
    Operations directorLeft the call after ten minutes without speaking; received a copy of the emailYesUnknown, though the warehouse manager expects the director to review the resultsAsk the warehouse manager what the director’s role is and what they will look for
    Finance contactReceived a copy of the email; hasn’t spokenYesUnknown. Finance approval is required at $26,000, but who gives it isn’t statedAsk who in finance approves, and what they need to see

    Questions about their process

    1. Who makes the final decision after the trial, and what will they want to see?
    2. Who in finance approves spending over $20,000, and how long does that usually take?
    3. Does anyone need to approve how incident footage is stored before the purchase goes ahead?

    What to noticeThe director said nothing, but the warehouse manager’s email suggests the director’s view matters. Marking the director’s role unknown prompts you to ask, without assuming a role or leaving the director out.

    The instruction behind this result, underlined in step 1“If someone only attended or received an email without replying, mark their role unknown and ask about it.”

Why it works

The table pairs each person’s role with what they said or did, so you can check the evidence for that role. It distinguishes confirmed roles from roles you still need to ask about. Attending a call or receiving an email doesn’t, by itself, tell you what part someone plays in the decision.

Check the answer

  • Every confirmed role rests on something the person said or did, or on a process rule.
  • People who only attended or received an email have an unknown role and a question to clarify it.
  • Your likeliest supporter’s row says what they agreed to do, not how keen they seem.

Follow up

When you learn more about their process:

Prompt 5 of 10

Check whether your offer really fits

Use it when you think“I want to call this a strong match, but one requirement is bugging me.”

Copy the prompt

You’ll need

  • Their must-haves and nice-to-haves
  • What our offer does, from our documents
  • Other ways to meet the requirement, if any

Pricing, discount limits, deal histories and requests for proposal (RFPs) are often confidential. Use an AI tool your company approves, share only what you’re allowed to, and replace names with roles.

Replace each [bracket] with your own material

Check whether our offer meets this customer’s requirements before I call it a match.

REQUIREMENTS
[Their must-haves and nice-to-haves]

OUR OFFER
[What our offer does, from our documents]

OTHER OPTIONS
[Other ways to meet the requirement, if any]

Make a table: requirement, must-have or nice-to-have, what our documents show, and status (met, partly met, not met or unknown). Count any feature planned for the future as not available today. For an alternative solution, say what it costs and who would support it; if our documents don’t say, list it as something to check. A met nice-to-have can’t make up for a missed must-have. If a must-have isn’t fully met, say so plainly and suggest how to raise it with the customer.

Give me: the table, the strongest case against our offer, then the next step.

See it work

With this prompt

Example written for this page

  1. YouThe prompt from step 1, filled in with:

    Operations lead at a 12-shop bakery chain: “We must keep taking card payments when the internet drops, and we need one sales report for all shops. A loyalty program would be nice.” Our docs: one account covers any number of shops, with combined sales reports, and loyalty is built in. Card payments work offline for up to four hours; longer offline use is planned for next year. Partner option: Overline’s 4G backup router keeps registers online if broadband fails.

  2. AIRequirement by requirement

    RequirementMust or niceWhat our documents showStatus
    Keep taking card payments when the internet dropsMust-haveOffline card payments for up to four hours; longer use is only planned for next yearPartly met: outages over four hours aren’t covered today
    One sales report for all 12 shopsMust-haveOne account covers any number of shops, with combined reportsMet
    Loyalty programNice-to-haveBuilt inMet

    Strongest case against our offer

    If the chain’s outages can last more than four hours, a must-have is only partly met, and a built-in loyalty program can’t make up for that. Overline’s router might close the gap, but our documents don’t give its price or say who supports it, so it can’t count yet.

    Next step

    Tell the operations lead plainly that card payments keep working offline for up to four hours, and ask how long their outages usually last. If outages can last longer, get Overline’s price and support terms before suggesting the router.

    What to noticeA quick yes would hide the four-hour limit, and a feature planned for next year doesn’t change today’s status. The next step asks the customer the one question that decides whether the gap matters.

    The instruction behind this result, underlined in step 1“Make a table: requirement, must-have or nice-to-have, what our documents show, and status (met, partly met, not met or unknown).”

Why it works

A status for every requirement makes the AI show where your offer meets the customer’s needs and where it falls short. The prompt keeps future features from counting as available today. It also asks for the cost and support arrangements of any alternative solution before treating it as an answer to a missing requirement.

Check the answer

  • Every status comes from your documents, and no planned future feature counts as available today.
  • An alternative solution’s cost and support are stated or listed to check, not assumed.
  • Any must-have that isn’t fully met is raised with the customer in the next step.

Follow up

When you’re ready to tell the customer about the gap:

Prompt 6 of 10

Check your offer’s value using real figures

Use it when you think“I need to show what the customer could gain without making up savings.”

Copy the prompt

You’ll need

  • Figures the customer gave you
  • Your offer, price and what it includes

Pricing, discount limits, deal histories and requests for proposal (RFPs) are often confidential. Use an AI tool your company approves, share only what you’re allowed to, and replace names with roles.

Treat the customer’s cost figures as confidential too, and leave out their company name if you can.

Replace each [bracket] with your own material

Help me compare our price with the benefits this customer could get, using their own figures.

THEIR FIGURES
[Figures the customer gave you]

OUR OFFER
[Your offer, price and what it includes]

Use only these figures, and say who gave each one; label estimates as estimates. Don’t add industry averages or typical savings. Show every calculation so I can check it. Where savings are unknown, calculate the benefit needed to cover our price. Keep the value of time saved separate from money gained through lower costs or extra profit. Turn missing figures into questions for the customer.

Give me: the figures and who gave them, the calculations, then the questions.

See it work

With this prompt

Example written for this page

  1. YouThe prompt from step 1, filled in with:

    Call with the practice manager at a dental group, October 13: “Our two billing staff spend about 60 hours a month fixing rejected insurance claims.” She puts their time at $30 an hour. “If this halved the rework, it would be an easy yes.” Our offer: ClaimSieve flags missing or mismatched codes before a claim is sent. $600 a month, plus a one-time $1,200 setup.

  2. AIFigures, calculations, questions

    Figures and who gave them

    • Time spent fixing rejected claims: about 60 hours a month (the practice manager’s estimate).
    • Cost of that time: $30 an hour (the practice manager’s estimate).
    • ClaimSieve: $600 a month, plus a one-time $1,200 setup (our offer).
    • Time ClaimSieve would save: not known. “If this halved the rework” is her what-if, not a measured result.

    Calculations

    • Rework today: 60 hours × $30 = $1,800 a month, if her estimate holds.
    • Monthly fee in staff time: $600 ÷ $30 = 20 hours.
    • First year: ($600 × 12) + $1,200 = $8,400, which is 280 hours of staff time ($8,400 ÷ $30), or 23⅓ hours a month.
    • At her $30-an-hour value, saving 23⅓ hours a month equals the first-year price in staff time, not necessarily cash savings; the threshold is 20 hours a month after that. Her hypothetical saving of half the time (30 hours) would be worth more than that price, but no one has measured it.
    • The financial return is still unknown: it depends on how those hours reduce costs or add profit after any extra costs.

    Questions for the practice manager

    1. How many of the 60 hours are spent on claims rejected for missing or mismatched codes, the errors ClaimSieve flags?
    2. Is the 60 hours tracked, or a rough estimate? Could you count last month’s rejected claims?
    3. Would saved hours cut overtime or other costs, add profit after any extra costs, or free your staff for other work?

    What to noticeThe answer shows how many hours saved would have a value equal to the price. It does not count that value as cash saved until the customer explains how those hours would reduce costs or add profit.

    The instruction behind this result, underlined in step 1“Keep the value of time saved separate from money gained through lower costs or extra profit.”

Why it works

Using only the customer’s figures and showing every calculation gives you a proposal you can check. If the savings are unknown, the prompt calculates how much benefit would be needed to cover your price. Time saved may let staff do other work, but it is a cash saving only if costs actually fall.

Check the answer

  • Every figure comes from the customer or your offer, says who gave it, and estimates are labeled.
  • Each calculation shows its work, and one you redo by hand gives the same result.
  • The answer separates the value of time saved from lower costs or added profit, and asks about missing benefits.

Follow up

When the customer answers your questions:

Prompt 7 of 10

Ask what “too expensive” means

Use it when you think“They said it’s too expensive, and I don’t actually know what they mean.”

Copy the prompt

You’ll need

  • What the buyer said, word for word
  • Your price and what you can change

Pricing, discount limits, deal histories and requests for proposal (RFPs) are often confidential. Use an AI tool your company approves, share only what you’re allowed to, and replace names with roles.

Replace each [bracket] with your own material

Help me work out why this buyer thinks our price is too high before I reply.

WHAT THEY SAID
[What the buyer said, word for word]

PRICE AND OPTIONS
[Your price and what you can change]

List up to three things the objection could mean, using only what the buyer has said, and quote the words that point to each. Write one neutral question that would tell these meanings apart. Then draft a short reply for each likely answer, using only what I can change, such as what the offer includes, the schedule or when payment is due. Don’t invent savings, competitor prices or discounts. Write the question and replies the way I’d say them to the buyer.

Give me: the possible meanings, the question, then a reply for each answer.

See it work

With this prompt

Example written for this page

  1. YouThe prompt from step 1, filled in with:

    The office manager at an accounting firm, earlier in the call: “We pay our current cleaners $2,900 a month, and they keep skipping the kitchen.” Then: “At $4,200 a month, you’re too expensive.” Our quote: five nights a week, kitchen and restrooms every night, $4,200 a month. I can offer three nights a week for $2,800, or $3,900 a month on a two-year contract. No other discounts.

  2. AIMeanings, question, replies

    What “too expensive” could mean

    1. The extra $1,300 a month over what they pay now may not seem worth it. Words that suggest this: “We pay our current cleaners $2,900 a month, and they keep skipping the kitchen.” This fits best, since the office manager raised the comparison, though what the $2,900 buys isn’t known.
    2. It’s over a budget they have to stay within. Nothing they have said points to a budget yet, but “too expensive” alone could mean it.

    Question to ask

    Is there a monthly budget you need to stay within, or is it more about whether paying more than you do now is worth it?

    Draft replies

    • A budget: “Thanks, that helps. What monthly figure do you need to stay within? Three nights a week is $2,800, or five nights is $3,900 a month on a two-year contract.”
    • Whether it’s worth it: “That’s fair. What does your current $2,900 cover, and how many nights? Ours is five nights a week, with the kitchen and restrooms cleaned every night.”

    What to noticeOne question separates a budget limit from doubts about value, so you answer the real concern. Each reply first asks for the fact it needs, and only the budget reply offers options, all within what you can change.

    The instruction behind this result, underlined in step 1“Write one neutral question that would tell these meanings apart.”

Why it works

Asking for the possible meanings, each with the buyer’s own words, shows why the AI reads the objection one way. You can check that reading before you answer. One question that separates the meanings, with a reply for each answer, lets you address the real concern and hold back offers until needed.

Check the answer

  • Each possible meaning quotes the buyer, or says plainly that nothing points to it yet.
  • The question doesn’t assume an answer, and the replies stay within what you can change.
  • No reply invents a saving, a competitor price or a discount.

Follow up

When the buyer answers your question:

Prompt 8 of 10

Rehearse a difficult sales conversation

Use it when you think“I freeze when a buyer pushes for a deadline we can’t meet, and I want to practice before a real call.”

Copy the prompt

You’ll need

  • The buyer and situation, without names
  • Your product facts and limits
  • What you want to practice

Pricing, discount limits, deal histories and requests for proposal (RFPs) are often confidential. Use an AI tool your company approves, share only what you’re allowed to, and replace names with roles.

Replace each [bracket] with your own material

Run a practice conversation in which you play the buyer and I play the seller.

SCENARIO
[The buyer and situation, without names]

OUR OFFER
[Your product facts and limits]

MY GOAL
[What you want to practice]

Stay consistent with the scenario. Push back, ask follow-up questions or show impatience as this buyer plausibly would, but don’t invent product facts, prices or competitor offers. Speak one turn at a time and wait for my reply; never write my lines. When I type “Review,” stop and assess each part of my goal in turn, quoting my words. Then give me one better line to try, and ask me to retry that moment.

Give me: the buyer’s first line, and nothing else.

See it work

With this prompt

Example written for this page

  1. YouThe prompt from step 1, filled in with:

    Scenario: the buyer, an events manager at a university, needs 200 printed tote bags for a conference on Friday. It’s Tuesday. Our limits: printing takes six business days after artwork approval, so Wednesday of next week at the earliest; we don’t offer rush printing. Plain bags can arrive Thursday. Goal: acknowledge her deadline, stay within what we can deliver, and ask a useful question before suggesting a next step.

  2. AIThe buyer opens

    Our conference starts Friday, and I need 200 printed tote bags there for the attendees. Can you do it?

    What to noticeThe AI plays only the buyer and stops after one line, so the next move is yours.

    The instruction behind this result, underlined in step 1“Speak one turn at a time and wait for my reply; never write my lines.”

  3. YouYour reply

    Printing takes six business days, so Friday isn’t possible.

    What to noticeThis scripted reply is accurate, but it rules Friday out without acknowledging why it matters, and asks nothing.

  4. AIThe buyer pushes back

    That doesn’t help me. I’ve got 200 people arriving Friday morning. Is there really nothing you can do?

    What to noticeThe buyer presses as a real one might but stays within the scenario: no rush service appears, and no rival quote is invented.

  5. YouYour reply, then “Review”

    Sorry. We could get plain bags to you by Thursday.

    Review

    What to noticeThe reply offers the plain bags, then “Review” stops the practice and asks for feedback.

  6. AIThe review

    Acknowledge her deadline

    Not yet. “Friday isn’t possible” names her date only to rule it out, and “Sorry” responds to her frustration, but neither reply shows you understand what Friday means for her 200 attendees.

    Stay within what we can deliver

    Yes. “Printing takes six business days” and “plain bags to you by Thursday” both match the scenario.

    Ask a useful question before suggesting a next step

    No. You offered plain bags before asking what the bags are for on Friday.

    One better line to try

    “I know Friday can’t move, with 200 people arriving. Printed bags would be ready Wednesday of next week at the earliest. What will the bags be used for at the conference? Then I can tell you what we can still get to you in time.”

    Try that moment again: how would you answer “Is there really nothing you can do?”

    What to noticeThe review takes each part of the goal in turn and quotes the actual replies. The better line acknowledges the deadline, states the limit and asks before offering anything.

Why it works

Asking the AI to play only the buyer and wait after each line makes you practice answering in your own words. The buyer’s pushback adds realistic pressure without invented facts. The review quotes your words for each part of your goal, so you see exactly what to change.

Check the answer

  • The AI speaks only as the buyer and waits for your reply each time.
  • The buyer pushes back, but every product fact, price and date matches your scenario.
  • The review covers each part of your goal and quotes your actual words.

Follow up

When you’re ready for a harder version:

Prompt 9 of 10

Check a request for proposal in sections

Use it when you think“This request for proposal is sixty pages, and I’m tempted to answer yes to everything.”

Copy the prompt

You’ll need

  • One numbered section of the proposal request
  • Our documents that answer it

Pricing, discount limits, deal histories and requests for proposal (RFPs) are often confidential. Use an AI tool your company approves, share only what you’re allowed to, and replace names with roles.

Requests for proposal and your own security reports often have confidentiality terms. Check those terms before pasting any part into an AI tool.

Replace each [bracket] with your own material

Help me check a request for proposal (RFP) against our documents, one section at a time. An RFP lists what a customer needs an offer to include.

RFP SECTION
[One numbered section of the proposal request]

OUR DOCUMENTS
[Our documents that answer it]

Make a table of every requirement in this section, splitting any that ask for two things. For each, give its number, whether it’s mandatory, its status (met, partly met, not met or unknown) and the line from our documents that shows it. A related feature doesn’t count as met. List deadlines and required formats separately, including any deadline for questions. Ask our team for missing evidence about our product; ask the customer to clarify requirements or accept an alternative for an unmet mandatory item.

Give me: the table, deadlines and formats, internal questions, customer questions, then the item numbers covered and requirement count.

See it work

With this prompt

Example written for this page

  1. YouThe prompt from step 1, filled in with:

    RFP 4, Security (all mandatory). 4.1 Staff use their city login through SAML, the required sign-in standard. 4.2 Keep system-activity records (audit logs) seven years; export on request. 4.3 Provide an independent security audit report from the last 12 months. Questions due October 30, 5 p.m. Eastern. Proposals due November 13, as a PDF. Our guide: “Supports SAML single sign-on. Audit logs are kept for 12 months and can be exported at any time.”

  2. AIOne section, checked

    RequirementMandatory?StatusWhat our documents say
    4.1 Staff sign in through the city’s single sign-on (SAML)YesMet“Supports SAML single sign-on.”
    4.2a Keep audit logs for seven yearsYesNot met“Audit logs are kept for 12 months…”
    4.2b Export audit logs on requestYesMet“…can be exported at any time.”
    4.3 Independent security audit report from the last 12 monthsYesUnknownNot mentioned; ask our security team

    4.2a is mandatory and not met; unless the city accepts another approach, the proposal may be ruled out. Ask by October 30, 5 p.m. Eastern.

    Deadlines and formats

    • Questions: October 30, 5 p.m. Eastern.
    • Proposals: November 13, as a PDF. No time of day is given.

    Questions for our team

    • 4.3: Ask our security team whether we have an independent security audit report from the last 12 months, and request it if available.

    Questions for the customer

    • 4.2: Would the requirement to keep logs for seven years be met if the city kept exported copies of the audit logs in its own storage?
    • What time on November 13 are proposals due, and in which time zone?

    Items covered

    Section 4: items 4.1, 4.2 and 4.3; four requirements after splitting 4.2.

    What to noticeSplitting requirement 4.2 shows that the product can export logs but cannot keep them for seven years. Ask your security team about the missing audit report; ask the customer only about their requirements or an acceptable alternative.

    The instruction behind this result, underlined in step 1“Make a table of every requirement in this section, splitting any that ask for two things.”

Why it works

Working one section at a time, with combined requirements split and each status backed by a quoted line, gives you a table you can check. AI tools are more likely to skip or merge items in long documents. A list of the item numbers covered lets you compare with the RFP’s own numbering and spot anything missing.

Check the answer

  • Every numbered item in the section has a row, and a split item’s parts keep its number, such as 4.2a and 4.2b.
  • Each “met” quotes evidence for the whole requirement; missing evidence about your product goes to your team.
  • The deadlines for questions and for proposals, with times and formats, are listed or flagged as missing.

Follow up

For each remaining section:

Prompt 10 of 10

Work out why a deal went quiet

Use it when you think“They stopped replying after the product demonstration, and I’ve already sent two follow-up messages.”

Copy the prompt

You’ll need

  • What happened so far, with dates
  • What you can offer or do next
  • Today’s date

Pricing, discount limits, deal histories and requests for proposal (RFPs) are often confidential. Use an AI tool your company approves, share only what you’re allowed to, and replace names with roles.

Replace each [bracket] with your own material

Help me work out why this deal has gone quiet and what to do next, without jumping to conclusions about the buyer’s motives.

HISTORY
[What happened so far, with dates]

MY OPTIONS
[What you can offer or do next]

TODAY
[Today’s date]

Start with a short timeline of what each side agreed to do and what actually happened. Keep the buyer’s own target dates apart from dates we both agreed. Give up to three possible explanations, each with the evidence for and against it, and don’t treat silence as a no. Recommend one next step from my options and say what it would tell us. If that step is a message, draft it in 80 words or fewer, with no pressure or invented urgency.

Give me: the timeline, the explanations, the next step, then the message.

See it work

With this prompt

Example written for this page

  1. YouThe prompt from step 1, filled in with:

    September 22: product demonstration for the HR director and new-hire setup lead. Director: “We’ll talk it over and get back to you by October 2.” September 23: sent requested pricing. October 6 and 13: follow-ups, no reply. October 15: setup lead asked if new hires can complete the process on phones; I said yes. I can offer a 20-minute call with our setup specialist or a customer reference. Today: October 19.

  2. AITimeline, reasons, draft message

    Timeline

    • September 22: product demonstration for the HR director and her new-hire setup lead. The director said they would talk it over and get back to us by October 2. That was her own target; no next meeting was agreed.
    • September 23: we sent the pricing she asked for, the day after the product demonstration.
    • October 2: her target passed without word.
    • October 6 and 13: two check-ins to the director, no reply.
    • October 15: the new-hire setup lead asked whether new hires can complete their setup on their phones. We answered the same day.

    Possible explanations

    1. Their discussion is taking longer than the director expected. For: the new-hire setup lead’s question on October 15, four days ago, shows someone is still looking at it. Against: the director hasn’t replied since the product demonstration.
    2. The new-hire setup lead is now doing the detailed review, and the director is waiting on it. For: the latest question came from the new-hire setup lead. Against: nothing says the director has handed it over, and the new-hire setup lead asked only one narrow question.
    3. Price is a concern. For: the director hasn’t replied since we sent the pricing. Against: no one has mentioned price, and the latest question was about how the product works.

    Next step

    Offer the director and the new-hire setup lead the 20-minute call with our new-hire setup specialist. It gives them a concrete next step, and whether they accept will show how active the evaluation is.

    Draft message

    Hello,

    The question about new-hire setup on phones made me think you may have others as you talk it over. Would a 20-minute call with our new-hire setup specialist help?

    If the timing has changed, just let me know.

    What to noticeOctober 2 was the director’s own target; both sides had not agreed to a next meeting. The missed date does not, by itself, show that the buyer has rejected the offer.

    The instruction behind this result, underlined in step 1“Keep the buyer’s own target dates apart from dates we both agreed.”

Why it works

The timeline distinguishes dates the buyer suggested from dates both sides agreed to. Comparing evidence for and against each explanation helps you see whether there are still signs of interest. The message then offers a useful next step without inventing a deadline.

Check the answer

  • The timeline separates the buyer’s own target dates from dates both sides agreed.
  • Each explanation has evidence for and against it, and silence isn’t read as a no.
  • The message offers something useful from your options, with no pressure or deadline.

Follow up

When the buyer replies:

Marketing

Prompt 1 of 10

Check the claims in your advertising

Use it when you think“My draft says we’re the best and customers love us, but I can’t prove either.”

Copy the prompt

You’ll need

  • Your draft advertising or promotional text
  • Product facts and offer terms, numbered

Customer quotes and interview notes can be personal data, and unreleased product facts or results may be confidential. Replace names with labels such as C1, and follow your organization’s rules on sharing with AI tools.

Replace each [bracket] with your own material

Check my advertising or promotional text against my facts, then revise it so it promises only what the facts support.

DRAFT
[Your draft advertising or promotional text]

FACTS
[Product facts and offer terms, numbered]

List every promise a buyer could take from the draft, including ones implied by words like “best” or “loved.” Next to each, give the fact number that supports it, or write UNSUPPORTED. Narrow each unsupported promise to what the facts support, or cut it, and keep the text persuasive. Keep each offer’s conditions next to the offer. If a buyer would need something the facts leave out, such as what happens after a free period, ask me instead of guessing.

Give me: the claim list, then the revised advertising or promotional text, then any questions.

See it work

With this prompt

Example written for this page

  1. YouThe prompt from step 1, filled in with:

    Draft: “The freshest coffee in the country, roasted to order and loved by thousands. Get your first bag free.”
    Facts: F1: every bag ships within 48 hours of roasting. F2: 2,300 active subscribers. F3: the first bag is free with a new subscription. F4: a subscription costs $16 a bag, shipped every two weeks.

  2. AIClaims, then the revised ad

    Claims

    • “The freshest coffee in the country”: UNSUPPORTED. None of the facts compares you with other roasters.
    • “Roasted to order”: UNSUPPORTED. F1 says every bag ships within 48 hours of roasting, not that it’s roasted after the order comes in.
    • “Loved by thousands”: “thousands” is backed by F2, 2,300 active subscribers. “Loved” is UNSUPPORTED: subscribing isn’t the same as loving it. Reviews or a survey could show whether subscribers do.
    • “Get your first bag free”: F3, but only with a new subscription, which the draft leaves out.

    Revised draft ad

    Coffee that ships within 48 hours of roasting. Start a subscription and your first bag is free; after that, it’s $16 a bag every two weeks. Join 2,300 subscribers.

    Questions for you

    • How do subscribers cancel, and can they cancel before the second bag ships? Buyers need this before they sign up, so it belongs next to the offer.
    • Is shipping charged on the free bag, or on the $16 bags?

    What to notice“Roasted to order” sounds like F1 but promises more: that each bag is roasted after the customer orders it. Matching each phrase to a numbered fact exposes the gap, and the revised ad keeps the real selling point, the 48 hours.

    The instruction behind this result, underlined in step 1“Next to each, give the fact number that supports it, or write UNSUPPORTED.”

Why it works

Asking the AI to list every promise against a numbered fact makes it show its evidence, so an unsupported word like “loved” stands out. The rule on offer conditions keeps the small print next to the offer, where a buyer will see it. US and UK advertising regulators (the FTC and the ASA) expect evidence for implied claims too, before the ad runs.

Check the answer

  • Every promise in the revised ad is supported by a fact number in the list, including ones a single word implies.
  • The offer’s conditions and the price afterward sit right next to the offer.
  • Where a claim was narrowed rather than cut, it still names a specific selling point, such as the 48 hours.

Follow up

When you need a shorter version for an email, ad or social post:

Prompt 2 of 10

Outline an article from your notes

Use it when you think“When AI drafts my blog posts, it adds statistics I can’t trace to any source.”

Copy the prompt

You’ll need

  • Article topic, intended readers and goal
  • Your notes and their sources

Customer quotes and interview notes can be personal data, and unreleased product facts or results may be confidential. Replace names with labels such as C1, and follow your organization’s rules on sharing with AI tools.

Replace each [bracket] with your own material

Outline an article from my notes, using only the facts and figures in them.

TOPIC
[Article topic, intended readers and goal]

NOTES
[Your notes and their sources]

Give a working title and four to six sections, each with the points it will make. After every number, statistic or quote, name the note it comes from, and keep details of who or what it describes, such as who was surveyed. Where a point needs evidence my notes don’t have, write NEEDS SOURCE and say what to look for; don’t fill the gap with a figure. Flag any figure in my notes that has no evidence behind it. Don’t mention studies, reports or experts that aren’t in my notes.

Give me: the outline, then a list of the sources I still need.

See it work

With this prompt

Example written for this page

  1. YouThe prompt from step 1, filled in with:

    Topic: how cafés can cut food waste. Reader: café owners. Goal: download our free waste-tracking sheet.
    Notes: our 2025 survey of 212 customer cafés found 61% throw away unsold pastries daily. A café owner we interviewed says baking in two smaller batches cut her pastry waste. A trade-show speaker said cafés waste about a third of their food.

  2. AIOutline with every figure traced

    Working title

    How to cut your café’s food waste, starting with pastries

    Sections

    1. The daily pastry problem: 61% of the 212 customer cafés we surveyed in 2025 throw away unsold pastries daily (2025 survey).
    2. How much food do cafés waste? NEEDS SOURCE: a published figure for food waste in cafés. Your notes give no evidence for the trade-show “about a third,” so I’ve left it out.
    3. Bake in smaller batches: one café owner says baking in two smaller batches cut her pastry waste (owner interview). NEEDS SOURCE: how much her waste fell; your note gives no figure.
    4. Start by measuring, with our free waste-tracking sheet. NEEDS SOURCE: evidence that tracking waste helps cafés cut it.

    Sources still needed

    • A published figure for food waste in cafés, to replace the trade-show estimate.
    • The owner’s before-and-after figures, and her permission to be quoted.
    • Evidence that tracking waste helps cafés cut it.

    What to noticeThis section needs a figure your notes don’t have, so the outline says what to look for instead of filling the gap. The trade-show figure stays out because your notes give no evidence for it.

    The instruction behind this result, underlined in step 1“Where a point needs evidence my notes don’t have, write NEEDS SOURCE and say what to look for; don’t fill the gap with a figure.”

Why it works

Asking the AI to name the note behind every figure shows where each number came from, so you can check it in a minute. When evidence is missing, models tend to fill the gap with a plausible statistic or source, and some are wrong or made up. Writing NEEDS SOURCE marks the gap instead.

Check the answer

  • Every number and quote names the note it came from.
  • No study, report or expert appears that isn’t in your notes.
  • Each figure says who or what it describes, such as “212 customer cafés we surveyed,” not all cafés.

Follow up

When you’ve found the missing sources:

Prompt 3 of 10

Turn a vague request into a campaign plan

Use it when you think“I’ve been asked for an ‘exciting campaign’ and nothing else.”

Copy the prompt

You’ll need

  • The request, word for word
  • Approved facts, places to publish and budget
  • Your organization’s style and brand rules

Customer quotes and interview notes can be personal data, and unreleased product facts or results may be confidential. Replace names with labels such as C1, and follow your organization’s rules on sharing with AI tools.

Replace each [bracket] with your own material

Help me turn a vague request into a plan for the people making the campaign.

REQUEST
[The request, word for word]

WHAT WE KNOW
[Approved facts, places to publish and budget]

BRAND
[Your organization’s style and brand rules]

First, find the missing decision that would most change the work, such as who it’s for or what they should do. Ask me one question about it and wait for my answer. If nothing important is missing, say so and go on.

Then write the plan: goal, how to measure success, intended audience, one message, supporting facts, materials to create and limits. Use only the approved facts. Label anything I haven’t confirmed as proposed, and point out any conflict between the request and the brand rules.

Give me: the question first. After my answer, the plan and one tempting idea that would conflict with it, with the reason.

See it work

With this prompt

Example written for this page

  1. YouThe prompt from step 1, filled in with:

    Request: “We need an exciting campaign for the new season.”
    What we know: a six-play theater season. Season subscriptions and single tickets go on sale September 1, but last season’s subscribers can renew and keep their seats until August 15. Where it will appear: email and Instagram. Budget: $4,000.
    Brand: warm and plain; no hype words such as “unmissable.”

  2. AIOne question first

    Before I write anything: is the main goal to get last season’s subscribers to renew by August 15, or to sell tickets to new audiences when sales open on September 1? The answer changes the audience, the timing, the message and how we measure success.

    What to noticeThe request mixes two jobs that need different campaigns. One question settles which job this is before any creative work starts.

    The instruction behind this result, underlined in step 1“First, find the missing decision that would most change the work, such as who it’s for or what they should do.”

  3. YouThe answer

    Renewals. We had 1,850 subscribers last season.

    What to noticeOne answer settles the audience, the deadline and what success means.

  4. AIThe campaign plan

    Campaign plan

    • Goal: get last season’s subscribers to renew by August 15.
    • Success measure (proposed): renewal rate, counted as renewals by August 15 ÷ 1,850. Tell me your usual renewal rate and I’ll propose a target.
    • Audience: last season’s 1,850 subscribers.
    • Message (proposed): “Renew by August 15 to keep your seats for the new season.”
    • Supporting facts: six plays; subscribers who renew by August 15 keep their seats.
    • Materials to create (proposed): a renewal email, a reminder a week before August 15, and Instagram posts.
    • Limits: warm, plain language with no hype words; a $4,000 budget.

    Conflict to resolve

    The request asks for “exciting,” but the brand rules out hype words. Let the plays themselves carry the excitement rather than adjectives.

    An idea that conflicts with the plan

    “Our most thrilling season ever. Renew today.” It breaks the no-hype rule and makes a “most ever” claim that nothing supports.

    What to noticeOnce the goal is clear, the plan can specify how to count renewals. That gives the team a way to judge results beyond whether the campaign feels exciting.

Why it works

The first question identifies the missing decision that would most affect the campaign. Once you answer, the AI can write a plan around that goal. Asking how success will be measured gives your team a way to judge the result.

Check the answer

  • Your answer to the first question affects who sees the campaign, what it says or what the team creates.
  • The plan states a goal and how success will be counted.
  • Unconfirmed items are labeled proposed, and every claim comes from your approved facts.

Follow up

When your team has a first campaign idea or draft:

Prompt 4 of 10

Develop campaign ideas for different needs

Use it when you think“I ask for three campaign ideas and get the same message with different wording.”

Copy the prompt

You’ll need

  • Customer quotes, labeled C1, C2…
  • Product facts, offer and goal

Customer quotes and interview notes can be personal data, and unreleased product facts or results may be confidential. Replace names with labels such as C1, and follow your organization’s rules on sharing with AI tools.

Replace each [bracket] with your own material

Give me up to three campaign ideas, each built on a different reason my audience might act.

CUSTOMER QUOTES
[Customer quotes, labeled C1, C2…]

FACTS AND GOAL
[Product facts, offer and goal]

Keep the offer and what you ask people to do the same; change only their reason for doing it. For each idea, give the reason in the customer’s terms, the quotes and facts it rests on, a headline and one supporting sentence. Use only my product facts, and label any assumption. Two quotes that give the same reason in different words support one idea, not two, and different wording alone is not a new idea. Don’t pad the list to reach three. For each idea, suggest a check tied to my goal, a rule agreed before testing for judging the result, and when the result would be too uncertain to decide. Ask for missing test information before choosing rules or how many people to include.

Give me: the ideas, each with its check.

See it work

With this prompt

Example written for this page

  1. YouThe prompt from step 1, filled in with:

    Parents: C1: “I need something for the two weeks after school ends.” C2: “He plays games all day. I’d love him to make one instead.” C3: “In his last online course, he got stuck and gave up.” C4: “Summer childcare is a nightmare once school’s out.”
    Facts: F1: camp runs June 22 to July 3, 9 a.m. to 3 p.m. F2: every camper codes and publishes a simple game. F3: one instructor for every eight campers.
    Goal: online bookings.

  2. AIThree reasons, three ideas

    1. Plan the start of summer break

    Reason: parents need a plan for the first two weeks after school ends (C1; C4 says the same in other words).
    Headline: “The first two weeks of summer break, planned.”
    Support: “Coding camp runs June 22 to July 3, 9 a.m. to 3 p.m.” (F1)
    Assumptions: local schools finish just before June 22, and the camp is in person. Hours of 9 a.m. to 3 p.m. may not cover a working day, so don’t call it childcare.
    Check: compare the dates with local school calendars. If schools finish later, drop this idea.

    2. Make a game, not just play one

    Reason: parents want game time to turn into making something (C2).
    Headline: “From playing games to making one.”
    Support: “Every camper codes and publishes a simple game.” (F2)
    Check: compare the three headlines by randomly dividing the same audience into three equal groups. Show each group one headline and send all groups to the same booking page. Measure booking rate: bookings ÷ people shown each ad. The plan is incomplete: before choosing how to judge the result, how many people to include or when to stop, I need your current booking rate, expected number of people seeing the test and smallest useful improvement. Don’t launch this comparison or choose a winner until that information is supplied and the plan is agreed; clicks alone cannot settle the booking goal.

    3. Help when they get stuck

    Reason: parents worry their child will get stuck and give up (C3).
    Headline: “Getting stuck is part of coding.”
    Support: “At camp, there’s one instructor for every eight campers.” (F3)
    Check: use the same booking-rate comparison only after its plan is complete and recorded before launch. If the result is too uncertain under that plan, say the test does not give a clear answer.

    What to noticeC4 sounds like a new reason but repeats C1’s, so it backs idea 1 instead of padding the list. Each idea that remains answers a different worry, which is what makes the three worth testing against each other.

    The instruction behind this result, underlined in step 1“Two quotes that give the same reason in different words support one idea, not two”

Why it works

Keeping the offer and requested action the same lets you compare the reasons people might respond. Linking every idea to customer quotes shows whether it addresses a different need or repeats an earlier idea. Measuring the result you actually want, such as bookings, keeps clicks or early interest from being treated as success.

Check the answer

  • Each idea gives a different reason to act, tied to at least one customer quote.
  • Every product detail in the proposed headlines and sentences comes from your facts, and any assumption is labeled.
  • Each test has rules agreed in advance, or asks for missing information before any test begins.

Follow up

When you have checked the facts or completed a planned test:

Prompt 5 of 10

Find the buyer questions your page misses

Use it when you think“The page reads well, yet potential customers keep emailing us the same questions.”

Copy the prompt

You’ll need

  • Your product or service page text, by section
  • Questions buyers have asked
  • Product facts and offer terms

Customer quotes and interview notes can be personal data, and unreleased product facts or results may be confidential. Replace names with labels such as C1, and follow your organization’s rules on sharing with AI tools.

Replace each [bracket] with your own material

Check whether my product or service page answers the questions buyers actually ask.

PAGE
[Your product or service page text, by section]

QUESTIONS
[Questions buyers have asked]

FACTS
[Product facts and offer terms]

For each question, quote the page wording that answers it and mark it answered, partly answered, unanswered or contradicted. Contradicted means the facts say otherwise. Then suggest up to three changes, each naming its question, the section it changes and the fact behind the new wording. Don’t add new sections; if a change needs a new line, say where it should go. If no fact answers a question, tell me what to find out.

Give me: a table of questions, then the changes.

See it work

With this prompt

Example written for this page

  1. YouThe prompt from step 1, filled in with:

    Page: Headline: “Happy dogs, happy owners.” Subhead: “Book a local walker in two taps. Free cancellation, any time.” Button: “Book your first walk”
    Questions: Q1: “Will it be the same walker every time?” Q2: “What if I cancel the same morning?” Q3: “Can the walker give my dog her medicine?”
    Facts: F1: weekly bookings keep the same walker. F2: canceling before 8 a.m. on the day is free; after that, the walk is charged in full.

  2. AIQuestions mapped to the page

    QuestionPage wordingStatus
    Q1: same walker every time?NoneUnanswered. F1: weekly bookings keep the same walker.
    Q2: canceling the same morning?“Free cancellation, any time.”Contradicted: F2 charges in full after 8 a.m.
    Q3: giving medicine?NoneUnanswered, and no fact covers it.

    Changes

    1. Q2, line below the headline: replace “Free cancellation, any time” with “Free cancellation until 8 a.m. on the day of the walk.” (F2)
    2. Q1, add a sentence after the cancellation terms below the headline: “Book weekly and keep the same walker.” (F1)

    To find out

    Q3: whether walkers can give medicine, before the page says anything about it.

    What to noticeThe page answers the cancellation question, but its answer conflicts with the facts. An owner who cancels at 9 a.m. and is then charged in full is likely to complain, so this change comes first.

    The instruction behind this result, underlined in step 1“For each question, quote the page wording that answers it and mark it answered, partly answered, unanswered or contradicted.”

Why it works

Asking for the questions buyers actually send makes the AI review the page against their needs, one question at a time. The status column shows at a glance which answers are missing or wrong. Tying each change to a fact, and listing what to find out, helps stop the fix from making a promise your facts don’t support.

Check the answer

  • Each question is marked answered, partly answered, unanswered or contradicted, with the page wording quoted.
  • Each change names its question, a section the page has (or where a new line goes) and its fact.
  • Questions no fact answers are listed as things to find out, not answered with a guess.

Follow up

Before you publish the revised page:

Prompt 6 of 10

Compare competitors using their own pages

Use it when you think“When I ask AI about competitors, it quotes prices and features I can’t find anywhere.”

Copy the prompt

You’ll need

  • Your product or pricing page
  • Competitors’ pages and the date copied
  • The claim you want to make

Customer quotes and interview notes can be personal data, and unreleased product facts or results may be confidential. Replace names with labels such as C1, and follow your organization’s rules on sharing with AI tools.

Replace each [bracket] with your own material

Compare my product with competitors using only the pages below, not what you remember about them.

OUR PAGE
[Your product or pricing page]

THEIR PAGES
[Competitors’ pages and the date copied]

CLAIM
[The claim you want to make]

Make a table with one row per feature or price and one column per page. Where a page doesn’t mention something, mark it as not stated rather than filling it in. Compare plans with the same features and limits, and flag different billing periods. Then say whether the pages support my claim, and suggest a version they do support. In any claim you suggest, keep fees next to the price, and say what was compared and when.

Give me: the table, then your verdict on the claim.

See it work

With this prompt

Example written for this page

  1. YouThe prompt from step 1, filled in with:

    Our page (Fernpay): $12 a month, unlimited invoices, 2.9% fee per card payment.
    Ottermark, copied October 12: $10 a month billed yearly, or $13 billed monthly; 50 invoices a month.
    Tallyhawk, copied October 12: free for 5 invoices a month; Pro is $15 a month for unlimited invoices.
    Claim: “The cheapest unlimited invoicing for freelancers.”

  2. AIComparison and verdict

    FeatureFernpay (yours)Ottermark (Oct 12)Tallyhawk (Oct 12)
    Price$12 a month; billing frequency not stated$10 a month billed yearly, or $13 billed monthlyFree plan; Pro $15 a month; billing frequency not stated
    InvoicesUnlimited50 a monthFree plan: 5 a month. Pro: unlimited
    Card fees2.9% per paymentNot statedNot stated

    A price quoted per month does not tell you whether you pay each month or several months at once. Your page and Tallyhawk’s do not state when payments are collected.

    Verdict

    The pages don’t support “the cheapest.”

    • They cover two competitors, and “the cheapest” claims the whole market.
    • Ottermark caps invoices at 50 a month, so only Tallyhawk Pro is a plan that also offers unlimited invoices.
    • Card fees aren’t on their pages, so total cost can’t be compared; only plan prices can.

    Possible advertising claims

    • “Unlimited invoices for $12 a month, plus 2.9% per card payment.” The fee sits next to the price because buyers pay it too.
    • Compare with Tallyhawk only once you know its card fees. If they’re the same as yours or higher: “Unlimited invoices for $12 a month, plus 2.9% per card payment. That’s $3 a month less than Tallyhawk Pro’s plan price (checked October 12).” First confirm that both prices are billed monthly.

    What to noticeNeither competitor’s page lists card fees, so only plan prices can be compared, and the comparison says so. Both suggested claims keep your own 2.9% fee next to your price.

    The instruction behind this result, underlined in step 1“Where a page doesn’t mention something, mark it as not stated rather than filling it in.”

Why it works

A table built only from the pages you paste shows where every entry came from, so you can check it in a minute. Marking missing information as not stated guards against a well-known AI habit: filling gaps with plausible details. US and UK ad regulators (the FTC and the ASA) expect truthful, clear comparisons; the FTC also advises keeping cost terms near the price.

Check the answer

  • Every table entry comes from a pasted page, and missing information is marked as not stated.
  • Price comparisons match plans with the same limits and billing period.
  • Any comparison the answer suggests says what was compared and when, and keeps fees next to the price.

Follow up

When a competitor changes its page:

Prompt 7 of 10

Find the words your customers actually use

Use it when you think“I have interview notes, and the AI keeps turning them into an invented description of a typical customer.”

Copy the prompt

You’ll need

  • Interview excerpts, labeled C1, C2…
  • Who you spoke to and how you chose them

Customer quotes and interview notes can be personal data, and unreleased product facts or results may be confidential. Replace names with labels such as C1, and follow your organization’s rules on sharing with AI tools.

Your interview consent form may not cover AI tools, so check it first. Also remove details that could identify someone, such as an employer or a street name.

Replace each [bracket] with your own material

Find language I can use in marketing from these customer interviews, without inventing a typical customer.

EXCERPTS
[Interview excerpts, labeled C1, C2…]

SAMPLE
[Who you spoke to and how you chose them]

Make a table with the exact phrase, who said it, their situation and the need it suggests. Quote word for word, and keep your interpretation in the need column. If a line is my note rather than their words, mark it as a note and don’t quote it. Keep different priorities visible instead of blending them into one customer. Don’t add age, job, motives or how common a need is unless an excerpt says so. Say what this sample can’t tell me.

Give me: the table, up to three of their phrases worth testing, and one interview question that could show your main interpretation is wrong.

See it work

With this prompt

Example written for this page

  1. YouThe prompt from step 1, filled in with:

    Excerpts from members of our car-sharing service:
    C1: “It’s cheaper than owning a car. That’s the whole point for me.”
    C2: “Saturday mornings are impossible. Everything’s booked by Thursday.”
    C3: “I’d happily pay more if there were a car on my street.”
    Sample: three downtown members who answered a newsletter invitation.

  2. AITheir words and possible needs

    Exact phraseWhoSituationNeed it suggests
    “It’s cheaper than owning a car. That’s the whole point for me.”C1Compares sharing with owningLower cost than owning, above all
    “Saturday mornings are impossible.”C2Wants a car on Saturday morningsWeekend availability
    “Everything’s booked by Thursday.”C2Booking for the weekendWeekend availability
    “I’d happily pay more if there were a car on my street.”C3No car on their streetA car close to home, even at a higher price

    C1 says cost is the whole point; C3 would pay more for a car nearby. They may prioritize these differently; neither quote shows that cost and closeness can’t matter to the same person.

    What this sample can’t tell you

    How common any need is, or what non-members and people outside downtown want. These are three members who chose to reply.

    Phrases worth testing

    • “Cheaper than owning a car” (C1)
    • “A car on my street” (C3)

    They speak to the two different needs, so test them against each other. Use either as a claim only if your facts back it.

    Interview question

    “Have you ever had to choose between a cheaper car farther away and a more expensive one nearby, and what mattered most in that choice?” I interpreted these quotes as suggesting that members may give cost and closeness different priorities. Answers about that choice could show this interpretation is wrong; merely mentioning both needs would not settle it. Test whether a message should stress lower prices, nearby cars or both.

    What to noticeThe quotes suggest different priorities, but don’t prove there are separate kinds of customer. The next question asks how customers choose when a lower price means traveling farther, so you can test what to emphasize in your message.

    The instruction behind this result, underlined in step 1“Keep different priorities visible instead of blending them into one customer.”

Why it works

Asking for each exact phrase and who said it makes the AI show its evidence before its interpretation. You can check every insight against a real excerpt. Keeping different priorities visible, and ruling out details nobody mentioned, guards against a summary that blends different people into one customer who fits no one.

Check the answer

  • Every quote matches an excerpt word for word and says who said it.
  • No age, job, motive or “most customers” appears unless an excerpt says so.
  • Different priorities stay visible without assuming people care about only one need.

Follow up

After your next round of interviews:

Prompt 8 of 10

Show why customers should choose you

Use it when you think“Everyone likes our slogan, but I can’t say who it’s for or why customers would switch to us.”

Copy the prompt

You’ll need

  • Your product message or slogan
  • Facts and customer quotes, numbered
  • What customers use today instead

Customer quotes and interview notes can be personal data, and unreleased product facts or results may be confidential. Replace names with labels such as C1, and follow your organization’s rules on sharing with AI tools.

Replace each [bracket] with your own material

Check the message I use to explain who our product is for and why customers should choose it.

MESSAGE
[Your product message or slogan]

EVIDENCE
[Facts and customer quotes, numbered]

ALTERNATIVES
[What customers use today instead]

The message should say who it’s for, what they use now and why they should choose us instead. Split my message into those three parts. If I give you a slogan, first write the full message it should suggest, then say whether the slogan fits. For each part, cite the fact or quote that supports it, or write UNSUPPORTED. Treat what a customer says they would do as a claim to test, not proof. If no evidence shows why someone would switch, say so instead of just describing the product.

Give me: the check, the full message in one or two sentences, whether the slogan fits (if supplied), and one test for the weakest part.

See it work

With this prompt

Example written for this page

  1. YouThe prompt from step 1, filled in with:

    Slogan for Saltbook: “Bookkeeping that tastes as good as your food.”
    Evidence: F1: restaurant owners get last week’s profit report every Monday. F2: we connect directly to the three most popular restaurant sales and payment systems. C1: “My accountant only looks at my numbers at tax time.” C2: “I’d pay double to know each week whether I made money.”
    Alternatives: an accountant once a year; owners’ own spreadsheets.

  2. AIThe message, then the slogan

    This is a slogan, so it doesn’t have to name all three parts. Here is the full message it should suggest, checked against your evidence.

    Check

    • Who it’s for: restaurant owners (F1, F2).
    • What they use now: an accountant who looks at the numbers once a year (C1, your alternatives). Owners who keep their own spreadsheets may need a different reason, such as F2’s connection to the restaurant’s sales system.
    • Why switch: last week’s profit report every Monday (F1), instead of a once-a-year look (C1).
    • C2’s “I’d pay double”: a stated willingness to pay. Treat it as something to test, not proof.

    Full message for your team

    For restaurant owners whose accountant looks at their numbers only at tax time, Saltbook sends last week’s profit report every Monday, so they know each week whether they made money.

    Does the slogan fit?

    Not yet. “Tastes as good as your food” hints at who it’s for, but not at what they use now or why they’d switch. Keep the wit if you like it, but the line should hint at the Monday report.

    Test for the weakest part

    The reason to switch rests on two quotes, one of them only a stated willingness to pay. Offer a paid first month at your real price to owners who use a once-a-year accountant. Decide in advance how many sign-ups would count as support, then count sign-ups, not compliments.

    What to noticeA memorable slogan may still give customers no reason to choose you. Writing the full message first shows what the slogan needs to suggest.

    The instruction behind this result, underlined in step 1“If I give you a slogan, first write the full message it should suggest, then say whether the slogan fits.”

Why it works

Asking what customers use today makes the comparison explicit: why choose your product instead? Writing that full message first gives you something to compare the slogan against. Treating “I’d pay double” as something to test preserves the distinction between what people say and what they actually do.

Check the answer

  • The full message names who it’s for, what they use now and why they should switch.
  • Each part cites a fact or quote, and stated intentions are marked as things to test.
  • The test counts what people do, such as sign-ups, against a number set in advance.

Follow up

When the test results come in:

Prompt 9 of 10

Explain results without guessing the cause

Use it when you think“The numbers went up after our change, and everyone’s calling it a win.”

Copy the prompt

You’ll need

  • The numbers, with dates and definitions
  • Changes to the page, visitors or measurement
  • The decision you need to make

Customer quotes and interview notes can be personal data, and unreleased product facts or results may be confidential. Replace names with labels such as C1, and follow your organization’s rules on sharing with AI tools.

Replace each [bracket] with your own material

Help me read these results before we decide anything.

RESULTS
[The numbers, with dates and definitions]

WHAT CHANGED
[Changes to the page, visitors or measurement]

DECISION
[The decision you need to make]

Show every rate as a division, such as 30 sales ÷ 1,000 visitors = 3%, so I can check it. Before comparing periods, check that they count the same thing in the same way. Keep what the numbers show apart from possible causes, and for each cause, say what would support or rule it out. A change that follows an edit doesn’t prove the edit caused it.

Give me: what the numbers show, the possible causes, what we can decide now and the next check.

See it work

With this prompt

Example written for this page

  1. YouThe prompt from step 1, filled in with:

    Our renovation quote page. Requests: 120 in March, 180 in April. Visitors: 4,000 in March, 7,500 in April.
    Changed on April 1: a new headline; paid search ads started; requests were counted at the thank-you page in March, but at the form’s first step in April.
    Decision: keep the new headline?

  2. AIRates first, then causes

    What the numbers show

    • First, the months don’t count the same thing. March counted forms submitted; April counts forms started. Every submitted form was started first, so April had at most 180 submissions.
    • Requests as counted: 120 → 180, up 50% (60 ÷ 120).
    • Visitors: 4,000 → 7,500, up 87.5% (3,500 ÷ 4,000).
    • Request rate: March 120 ÷ 4,000 = 3.0%. April 180 ÷ 7,500 = 2.4%, the highest possible rate of completed requests.

    Possible causes of the rise

    • More visitors: traffic rose 87.5% in the month the ads started. Supported if most of the extra requests came from ad visitors.
    • The change in how requests were counted: started forms include people who never submit. April’s count of submitted forms would settle it.
    • The season: renovation requests may rise in spring anyway. Last year’s March and April would show it.
    • The new headline: possible, but the rate fell from 3.0% to at most 2.4%. A test that randomly shows the two headlines to comparable groups of visitors could separate the headline’s effect from other changes.

    What we can decide now

    Don’t report the headline as a win. Keeping or reverting it is a judgment call for now; nothing here shows it helped or hurt.

    Next check

    Get April’s submitted forms from your form tool or inbox, then recalculate the rate.

    What to noticeThe count rose from 120 to 180, but visitors increased faster. April’s 2.4% counts forms started, so the rate of completed requests could only be that high or lower.

    The instruction behind this result, underlined in step 1“Show every rate as a division, such as 30 sales ÷ 1,000 visitors = 3%, so I can check it.”

Why it works

Asking for every rate as a division makes the arithmetic visible, so you can check it in seconds. Checking that both periods count the same thing helps catch a change in who visits or how results are counted before a bigger number is called a win. Listing each cause with what would rule it out tells you what to check next.

Check the answer

  • Every rate is shown as a division you can redo in a calculator.
  • Both periods count the same thing, or the answer says plainly that they don’t.
  • Each possible cause comes with what would support it or rule it out.

Follow up

When you have comparable numbers:

Prompt 10 of 10

Plan a test with one change at a time

Use it when you think“We tested two versions, changed three things, and now nobody knows which change made the difference.”

Copy the prompt

You’ll need

  • Current version and proposed change
  • Who sees it and the action you want
  • Weekly visitors and current action rate

Customer quotes and interview notes can be personal data, and unreleased product facts or results may be confidential. Replace names with labels such as C1, and follow your organization’s rules on sharing with AI tools.

Replace each [bracket] with your own material

Plan a test comparing two versions (an A/B test) that differ in one thing.

VERSIONS
[Current version and proposed change]

AUDIENCE AND GOAL
[Who sees it and the action you want]

VISITORS AND RATE
[Weekly visitors and current action rate]

If the versions differ in several ways, keep my main idea and propose two that differ only in that. Define who is included, how visitors are divided, and the rate: completed actions ÷ visitors. Add one measure of possible harm, such as lower profit per visitor. Set how many visitors each version needs and when to stop before launch. Show the inputs I should check in a sample-size calculator, which estimates how many visitors a test needs. If I can’t reach that number in about a month, say so. State assumptions, including the smallest useful improvement, so I can change them. Don’t promise a winner.

Give me: a short test plan.

See it work

With this prompt

Example written for this page

  1. YouThe prompt from step 1, filled in with:

    Plant shop product pages.
    Current: headline “Healthy plants, delivered” and button “Add to cart”
    Proposed: a countdown timer, headline “20% off your first order” and button “Shop the sale”
    Audience: first-time visitors. Goal: purchases.
    Traffic: about 3,000 first-time visitors a week; 2% buy.

  2. AIOne change, planned in advance

    • One change: the proposal changes the offer, headline and button, and adds a timer. Assuming the discount is your main idea, test it alone: A is the current page, and B adds “20% off your first order” under the current headline, with the same button and no timer.
    • Who’s included: first-time visitors to product pages, excluding staff.
    • Split: assign visitors at random, half to each version, and keep returning visitors on the same one.
    • Main outcome: purchase rate = purchases ÷ visitors, for each version.
    • Measure of possible harm: gross profit per visitor = (revenue − product costs) ÷ visitors, because 20% off can raise sales and revenue while cutting profit.
    • Number of visitors needed (sample size): about 3,800 visitors per version, assuming the smallest rise worth acting on is 1 percentage point (2% to 3%, 50% more than the starting rate). Other calculator inputs: starting rate 2%; a two-sided test, which checks for either a rise or a fall; 5% significance level (some calculators show 95% confidence), 80% power. That means a 5% chance of a false alarm if the discount changes nothing, and an 80% chance of catching a real 1-point rise. Calculators differ by method, so 3,300 to 4,000 is in line.
    • Stopping point: 7,600 visitors ÷ 3,000 a week ≈ 2.5 weeks. Run three full weeks, so weekdays and weekends are covered evenly, and don’t stop early. A rise to only 2.5% would need about 13,800 per version by the same method, roughly nine weeks: too long for your traffic.
    • No winner promised: if the test can’t reliably distinguish the rates, call the discount unproven, not a failure.
    • If you only need to know whether the whole new page wins, test it as one bundle first, then its parts.

    To redo the plan, tell me how much gross profit you make on each sale (a rise to 3% may not pay for 20% off), whether the timer is your main idea, or a smaller rise worth acting on.

    What to noticeVersion B changes only the discount offer, so the comparison tests that change. The button wording and timer can each be tested later.

    The instruction behind this result, underlined in step 1“If the versions differ in several ways, keep my main idea and propose two that differ only in that.”

Why it works

The current action rate and expected visitor count help determine how many people the test needs and how long it will take. Setting those requirements before launch makes the plan checkable. Changing one thing helps identify its effect, while a separate harm measure checks whether an improvement, such as more sales, comes with a cost.

Check the answer

  • The two versions differ in exactly one thing.
  • The desired result and possible harm are each measured, for example purchases or profit divided by visitors.
  • The visitor count and stopping point are set before launch, with inputs you can check in a sample-size calculator.

Follow up

When the test has run its planned length:

Human Resources

Prompt 1 of 10

Tie each job ad requirement to the work

Use it when you think“The draft ad asks for a degree and 10 years of experience, and no one can tell me why.”

Copy the prompt

You’ll need

  • Your draft job ad
  • The main tasks of the job
  • Requirements you’ve confirmed, and why

Use only an AI tool your employer has approved for HR data, and share the minimum. In a small team, a date or job title can identify someone even without a name.

Replace each [bracket] with your own material

Check this job ad so that every requirement is tied to the work.

JOB AD
[Your draft job ad]

MAIN TASKS
[The main tasks of the job]

CONFIRMED REQUIREMENTS
[Requirements you’ve confirmed, and why]

For each requirement in the ad, name the task it supports. Label it “keep” (a task needs it, or it’s confirmed), “ask the hiring manager” (no reason given) or “reword” (describes the person, not the work). Flag words, such as “digital native” or “recent graduate,” that could discourage applicants because of age, sex, race, religion, disability, nationality or another characteristic protected by law, for review before posting. Don’t add requirements, and don’t say whether the ad is lawful. Keep pay and benefits exactly as written.

Give me: a table of requirements, then a revised ad with open points marked.

See it work

With this prompt

Example written for this page

  1. YouThe prompt from step 1, filled in with:

    Draft ad: “Scheduling coordinator. Join our young, energetic team. You’re a digital native with a bachelor’s degree and 10+ years of scheduling experience. Weekend availability required. $52,000–$58,000 a year, plus health insurance.” Main tasks: build weekly shift schedules in our scheduling software, resolve shift conflicts, explain changes to staff. Confirmed: weekend availability, because schedules go out on Sundays.

  2. AIRequirements, then the ad

    Requirements

    In the adTask it supportsLabel
    “Young, energetic team”None; it describes people’s age, not the workReword; flag for review (age)
    “Digital native”Using scheduling software, but the phrase describes a generation, not a skillReword; flag for review (age)
    Bachelor’s degreeNone of the listed tasksAsk the hiring manager
    10+ years of scheduling experienceScheduling, but no task explains why 10 yearsAsk the hiring manager; flag for review (can exclude younger applicants)
    Weekend availabilitySchedules go out on SundaysKeep (confirmed)

    Draft revised ad

    Scheduling coordinator. You’ll build weekly shift schedules in our scheduling software, resolve shift conflicts and explain changes to staff. Weekend availability is required, because schedules go out on Sundays. $52,000–$58,000 a year, plus health insurance.

    Still needs the hiring manager’s decision: bachelor’s degree; 10+ years of scheduling experience.

    What to notice“Digital native” sounds like a skill, but it describes when someone grew up. Asking which task each requirement serves exposes it, and the flag marks it for review as a possible age-discrimination risk.

    The instruction behind this result, underlined in step 1“For each requirement in the ad, name the task it supports.”

Why it works

Naming the task behind each requirement shows which ones have no stated reason, so you know what to ask the hiring manager. The flag catches risky wording: “digital native” can discourage older applicants, and US federal law protects applicants aged 40 and over. An unneeded degree requirement can also be unlawful if it screens out more people of one race or sex, which US law calls disparate impact. UK and EU law calls it indirect discrimination and protects all ages, so 10 years of experience can be a risk there too.

Check the answer

  • Every requirement in the ad is in the table with its task, if any, and a label: keep, ask the hiring manager or reword.
  • Potentially discriminatory wording, such as “digital native,” is flagged for review.
  • Pay and benefits appear in the revised ad exactly as written, and confirmed requirements are kept.

Follow up

When the hiring manager answers your questions:

Prompt 2 of 10

Answer a question from the policy’s wording

Use it when you think“An employee asks if they’re entitled to something, and I want to answer today without getting it wrong.”

Copy the prompt

You’ll need

  • The employee’s question, without names
  • The policy, with its version and date
  • What you know about their situation

Use only an AI tool your employer has approved for HR data, and share the minimum. In a small team, a date or job title can identify someone even without a name.

If the question involves health, disability or a death in the family, describe it in general terms and leave out medical details.

Replace each [bracket] with your own material

Help me answer an employee’s question using only our policy and the facts below.

QUESTION
[The employee’s question, without names]

POLICY
[The policy, with its version and date]

FACTS
[What you know about their situation]

Quote the policy wording behind each point, including conditions and exceptions. If the policy doesn’t settle something, show which words leave it open and what the person responsible for the policy must decide. Don’t say what the law requires, but flag where a law might give the employee more than the policy, such as on leave, pay, hours or disability, for me to check, with legal advice if needed. Then write a warm reply of 100 words or fewer that answers what the policy settles and says what we’ll confirm.

Give me: the quoted policy wording and unanswered points, then the reply.

See it work

With this prompt

Example written for this page

  1. YouThe prompt from step 1, filled in with:

    Question: “My grandmother died on Sunday, and the funeral is on Friday. Can I take bereavement leave for the whole week?” Policy (version 4, January 2026): “Employees may take up to 3 days of paid bereavement leave for the death of an immediate family member, such as a spouse, partner, child, parent or sibling. Further time off may be taken as vacation or unpaid leave, with manager approval.” Facts: full-time; 6 vacation days left.

  2. AIPolicy wording, then the reply

    Policy wording

    • Paid bereavement leave: “up to 3 days of paid bereavement leave for the death of an immediate family member.”
    • More time off: “Further time off may be taken as vacation or unpaid leave, with manager approval.”

    Unanswered points

    • Whether a grandparent counts. The policy names family members “such as a spouse, partner, child, parent or sibling,” which gives examples, not a full list. The person responsible for the policy must decide.
    • If “the whole week” means 5 working days and a grandparent counts, that’s 3 paid days plus 2 days of vacation or unpaid leave. Those 2 days need manager approval, and the 6 vacation days left would cover them. If a grandparent doesn’t count, this policy doesn’t cover the time off, so check your vacation rules.
    • For you to check, with legal advice if needed: where this employee works, a law may give bereavement leave, or define family more widely than this policy. Check before refusing any days.

    Draft reply to the employee

    I’m very sorry to hear about your grandmother. Our policy gives up to 3 days of paid bereavement leave for an immediate family member, and I’m checking whether that includes grandparents. I’ll confirm as soon as I can. If you need more time, you can ask your manager about using vacation (you have 6 days left) or unpaid leave.

    What to notice“Such as” introduces examples, so the policy neither includes nor excludes grandparents explicitly. Quoting those words shows exactly what you need the person responsible for the policy to clarify.

    The instruction behind this result, underlined in step 1“If the policy doesn’t settle something, show which words leave it open and what the person responsible for the policy must decide.”

Why it works

Quoting the policy wording behind each point shows which parts answer the question and which leave room for interpretation. You can then ask the person responsible for the policy to settle the unanswered points. Questions about leave, pay or disability also need a legal check, because a law may give an employee more than the policy says.

Check the answer

  • Each point quotes the policy wording it relies on, including conditions and exceptions.
  • Open points, and places where a law may give more, are flagged instead of answered.
  • The reply promises nothing the policy doesn’t allow and mentions any approval it requires.

Follow up

When the person responsible for the policy settles the unanswered point:

Prompt 3 of 10

Check each judgment in a performance review

Use it when you think“A manager’s draft review says ‘reliable’ and ‘poor attitude,’ with no dated example of either.”

Copy the prompt

You’ll need

  • The draft review, with names removed
  • Dated notes from the review period
  • Goals or rating definitions, if any

Use only an AI tool your employer has approved for HR data, and share the minimum. In a small team, a date or job title can identify someone even without a name.

Leave out notes about sickness, health or personal circumstances, and keep the rating decision with the manager.

Replace each [bracket] with your own material

Check whether the evidence supports each judgment in this draft performance review.

DRAFT REVIEW
[The draft review, with names removed]

DATED NOTES
[Dated notes from the review period]

GOALS AND RATINGS
[Goals or rating definitions, if any]

For each judgment word or phrase, such as “reliable,” list the dated notes that support it, contradict it or leave it unclear. Say when a judgment rests on a single event. Name any part of the period with no notes, and say that missing notes are not evidence of poor work. If a note could be read either way, say what to ask the manager instead of choosing. Don’t suggest or change the rating.

Give me: a table of judgments and evidence, then the gaps, then questions for the manager.

See it work

With this prompt

Example written for this page

  1. YouThe prompt from step 1, filled in with:

    Draft review, January–June: “Reliable. Poor attitude in team meetings. Meets expectations.” Rating definition: “Meets expectations: delivers agreed work on time and to standard.” Notes: February 12, payroll processing finished on time. March 9, payroll processing finished on time. April 16, corrected an error in the payroll file. May 14, questioned the proposed reporting change in the team meeting. June 30, payroll processing finished two days late.

  2. AIEvidence, then questions

    Judgments and evidence

    JudgmentSupportsContradictsUnclear
    “Reliable”February 12 and March 9: payroll processing finished on timeJune 30: payroll processing finished two days lateApril 16: the note doesn’t say whose error was corrected
    “Poor attitude in team meetings”NoneNoneMay 14: questioned a proposed change. Rests on this single event, and questioning a change isn’t poor attitude in itself
    “Meets expectations”February 12 and March 9: on timeJune 30: two days lateApril 16: whose error was corrected? Nothing else records whether the work was “to standard”

    Gaps

    There are no notes for January. Missing notes are not evidence of poor work.

    Questions for the manager

    1. On April 16, whose error was corrected?
    2. Apart from May 14, what did you see in team meetings that led to “poor attitude”?
    3. Were there other payroll processing deadlines from January to June, and were they on time?

    What to notice“Poor attitude in team meetings” sounds like a pattern, but the only meeting note says the person questioned a proposal. Seeing that, the manager can add real examples or drop the phrase.

    The instruction behind this result, underlined in step 1“Say when a judgment rests on a single event.”

Why it works

Listing the dated notes behind each judgment shows which phrases rest on one event or none. That matters because research on ratings finds that judgments of someone’s traits often say more about the person giving the rating than about the employee. Naming gaps in the period keeps missing notes from being read as poor work.

Check the answer

  • Every judgment in the draft appears in the table with dated notes, or is marked as having none.
  • Notes that could be read either way become questions for the manager, not evidence.
  • The answer doesn’t suggest, change or defend the rating.

Follow up

When the manager answers your questions:

Prompt 4 of 10

Help a manager plan a difficult conversation

Use it when you think“A manager says someone ‘doesn’t care anymore’ and wants my backing.”

Copy the prompt

You’ll need

  • What happened, with dates and no names
  • The expectation and how it was shared
  • What the manager wants to achieve

Use only an AI tool your employer has approved for HR data, and share the minimum. In a small team, a date or job title can identify someone even without a name.

Leave out health and personal details, even if the manager mentions them, and keep any formal step, such as a warning, within your organization’s process.

Replace each [bracket] with your own material

Help me prepare a manager for a difficult conversation with someone on their team.

WHAT HAPPENED
[What happened, with dates and no names]

EXPECTATION
[The expectation and how it was shared]

GOAL
[What the manager wants to achieve]

Separate what was observed from the manager’s interpretation. If my notes don’t show the expectation was made clear to the person, say so. List at least three explanations that fit the same facts, including ones outside the person’s control. Write an opening that states the facts and asks for their view, without deciding why it happened. Plan the order of the conversation, with a pause for HR advice if the person mentions health, disability or caregiving responsibilities.

Give me: the facts and the manager’s interpretation, the explanations, the opening, then the order of the conversation.

See it work

With this prompt

Example written for this page

  1. YouThe prompt from step 1, filled in with:

    Notes: the weekly sales report is due Mondays at 10 a.m. On September 28 it arrived at 3:40 p.m., and on October 5 at 2:15 p.m. It is built from figures the finance team sends. Manager: “She doesn’t care anymore. Everyone knows the deadline.” Goal: get the reports in on time without starting a formal process.

  2. AIPlan for the conversation

    What was observed

    • The report is due Mondays at 10 a.m. It arrived at 3:40 p.m. on September 28 and at 2:15 p.m. on October 5.
    • It is built from the finance team’s figures.

    The manager’s interpretation

    • “She doesn’t care anymore.” The notes show two late reports, not a reason.
    • “Everyone knows the deadline.” Nothing shows the 10 a.m. deadline was made clear to her, so ask the manager how it was shared.

    Explanations that fit the same facts

    1. The finance figures may have arrived late on those Mondays.
    2. She may have understood the deadline differently, for example as end of day.
    3. Her workload or priorities may have changed in late September.
    4. She may be less engaged, as the manager believes.

    Opening for the manager

    “The sales report came in at 3:40 p.m. on September 28 and at 2:15 p.m. on October 5. It’s due at 10 a.m. on Mondays, and I want to make sure we’re working to the same deadline. Can you walk me through how those two Mondays went?”

    Order of the conversation

    1. Ask for her account and listen, including when the finance figures arrived.
    2. Agree on the deadline, and on what to do if the figures arrive late.
    3. Ask what would help her meet it.
    4. Agree on one next step and a date to check in.

    If she mentions health, disability or caregiving responsibilities, pause and get HR advice before going further. Don’t ask for medical details.

    What to noticeThe notes mention the finance figures but not when they arrived. Listing explanations that fit the same facts turns that detail into the first thing to ask, before anyone decides she doesn’t care.

    The instruction behind this result, underlined in step 1“List at least three explanations that fit the same facts, including ones outside the person’s control.”

Why it works

Asking how the expectation was shared, and for other explanations that fit the facts, tests the manager’s view before the meeting. The manager then walks in with a question instead of a verdict. That gives the conversation a chance to uncover a cause like late figures before anyone reaches for a warning.

Check the answer

  • The manager’s interpretation is kept apart from what was observed.
  • At least one explanation is outside the person’s control, and the opening asks for their view.
  • The answer says whether the notes show the expectation was made clear, rather than assuming it.

Follow up

After the conversation, to write it up:

Prompt 5 of 10

Write interview questions you can score

Use it when you think“Every interviewer asks their own questions, and afterward we argue about gut feel.”

Copy the prompt

You’ll need

  • The job’s main tasks and skills
  • A real situation from the job, if any
  • Job seniority and interview length

Use only an AI tool your employer has approved for HR data, and share the minimum. In a small team, a date or job title can identify someone even without a name.

Replace each [bracket] with your own material

Draft one interview question per skill, with a scoring guide every interviewer applies the same way.

TASKS AND SKILLS
[The job’s main tasks and skills]

REAL SITUATION
[A real situation from the job, if any]

SENIORITY AND LENGTH
[Job seniority and interview length]

Base each question on a real job situation, and mark any detail you invent for the hiring team to check. Make each question need a judgment call, not recall of a rule. Add one planned follow-up question (a probe) that doesn’t hint at the answer. Give the condition for asking it, so every candidate gets it the same way, and say how to score an answer that improves only after it. Describe a weak, an acceptable and a strong answer by what the answer shows, not by confidence, accent, background or similarity to the interviewer. Match the questions to the job’s seniority, and say if they won’t fit the interview length. Remind me that the hiring team must approve the guide before use.

Give me: each question with its probe, then its scoring guide.

See it work

With this prompt

Example written for this page

  1. YouThe prompt from step 1, filled in with:

    Skill: prioritizing competing requests. Role: entry-level operations assistant; 30-minute interview. Real situation: at 9 a.m., three requests arrive. One store’s card reader is down, so it can take cash only; sorting it out takes about two hours. The manager wants the weekly report by noon, which takes about two hours. A customer needs an invoice copy by 5 p.m. for an audit, which takes 10 minutes.

  2. AIQuestion and scoring guide

    Question

    “It’s 9 a.m., and three requests arrive at once. One store’s card reader is down, so it can take cash only; sorting it out takes about two hours. Your manager wants the weekly report by noon, which also takes about two hours. A customer needs an invoice copy by 5 p.m. for an audit, which takes 10 minutes. What do you do, in what order, and why?”

    Probe

    If the answer gives an order without reasons, ask: “What made you choose that order?” Ask it only then, in the same words, for every candidate.

    Scoring guide

    • Weak: gives an order without reasons, even after the probe.
    • Acceptable: gives a sensible order with reasons, such as the card reader first because the store is losing card sales, but doesn’t notice the conflict with the report.
    • Strong: notices that the card reader and the report can’t both be done by noon, chooses with a reason, and says they’d tell the manager or ask for help before noon. Fits the invoice in before 5 p.m.

    Score what the answer shows by the end, except that a conflict spotted only after the probe scores Acceptable, not Strong.

    Doesn’t count as evidence: confidence, accent, background or similarity to the interviewer. Nor do general claims such as “I’m very organized.”

    Before use

    Every detail comes from the situation you gave. The question needs no specialist knowledge, which suits an entry-level role, and as a single question it fits easily in 30 minutes. The hiring team must approve the question and the guide before use.

    What to noticeNo rule settles the order, so candidates must weigh the requests themselves. The guide says what spotting the conflict is worth unprompted and after the probe, so two interviewers score it the same way.

    The instruction behind this result, underlined in step 1“Describe a weak, an acceptable and a strong answer by what the answer shows, not by confidence, accent, background or similarity to the interviewer.”

Why it works

A real situation from the job lets the AI build a question that needs judgment, not one a well-prepared candidate can answer from a script. Describing each level by what the answer shows gives every interviewer the same yardstick. Deciding in advance how to score an answer the probe improves keeps scores based on evidence rather than confidence or how well the interviewer and candidate get along.

Check the answer

  • The question needs a judgment call; there’s no rule a candidate could simply recite.
  • The probe has one condition, doesn’t hint at the answer, and the guide says how to score answers it improves.
  • Each level describes what the answer shows, not confidence, accent, background or similarity to the interviewer.

Follow up

If you try the question on colleagues and your scores differ:

Prompt 6 of 10

Prepare a new hire to do the job

Use it when you think“Managers tell me new hires finish our introductory checklist and still can’t do the job.”

Copy the prompt

You’ll need

  • Your new-hire checklist
  • What the new hire must be able to do
  • Who can review their work, and when

Use only an AI tool your employer has approved for HR data, and share the minimum. In a small team, a date or job title can identify someone even without a name.

Replace each [bracket] with your own material

Turn this new-hire checklist into a plan that shows when a new hire can do the job, not just that they finished each step.

CHECKLIST
[Your new-hire checklist]

JOB TASKS
[What the new hire must be able to do]

SUPPORT
[Who can review their work, and when]

Copy every legal, safety, payroll and IT step into a separate “Required steps” list, unchanged and with its timing. For each job task, add a practice task that produces work someone reviews, and say what a good result looks like; mark standards I didn’t give as proposed. Use only the reviewers and time in my support list. If an item seems unnecessary, or a resource is missing, ask the manager instead of removing or inventing it.

Give me: the required steps, then a table of practice tasks and checks, then questions for the manager.

See it work

With this prompt

Example written for this page

  1. YouThe prompt from step 1, filled in with:

    Checklist for a new clerk who processes supplier payments. Day 1: Form I-9 and payroll forms; safety orientation; laptop and finance system login; read the invoice guide. Day 2: observe a senior clerk at work. Day 3: start processing real invoices. Outcome: enters supplier invoices accurately and catches duplicates before payment. Support: the person leading supplier payments can review practice work for one hour on day 2.

  2. AIRequired steps and practice

    Required steps (unchanged)

    • Day 1: Form I-9 and payroll forms
    • Day 1: safety orientation
    • Day 1: laptop and finance system login

    Practice tasks and checks

    WhenTaskReviewerGood result (proposed)
    Day 1Read the invoice guideNonePreparation only; it doesn’t show readiness
    Day 2After observing a senior clerk at work, enter a set of past invoices with one duplicate addedPerson leading supplier payments, one hourEvery entry matches its invoice, and the duplicate is flagged before payment
    Day 3Start processing real invoicesNone listedOnly once the day 2 batch is right (see question 2)

    Questions for the manager

    1. Is there a set of past invoices, or a training copy of the finance system, to practice on? Who can prepare it with one duplicate added? Without these, the day 2 task can’t happen.
    2. If the practice entries aren’t right by the end of day 2, should the new clerk still start real invoices on day 3?

    What to noticeA reviewed practice batch shows whether the new clerk can catch a duplicate before handling real invoices; reading and observing a colleague only show that steps happened. The day 1 legal and safety steps stay exactly as they were.

    The instruction behind this result, underlined in step 1“For each job task, add a practice task that produces work someone reviews, and say what a good result looks like”

Why it works

A reviewed piece of real work for each job task shows whether the new hire can do the job, not just whether they finished the steps. Keeping legal, safety, payroll and IT steps in a separate list protects them during the rewrite, so checks such as safety training stay in place. So does Form I-9, the US form that confirms a new hire’s identity and eligibility to work.

Check the answer

  • Every legal, safety, payroll and IT step from your checklist is in the required list, with its timing.
  • Each job task has a practice task whose result someone reviews, not just an activity to finish.
  • Only the reviewers and time you listed are used, and standards you didn’t give are marked proposed.

Follow up

After the review, when you know how the practice went:

Prompt 7 of 10

Find themes in employee survey comments

Use it when you think“I have 300 survey comments, and leadership wants the themes by Friday.”

Copy the prompt

You’ll need

  • The comments, each with a number
  • The survey question and who answered
  • What senior managers need to decide

Use only an AI tool your employer has approved for HR data, and share the minimum. In a small team, a date or job title can identify someone even without a name.

Paste comments only if what respondents were told about confidentiality allows it. Remove details, such as team or role, that could show who wrote them.

Replace each [bracket] with your own material

Find the themes in these employee survey comments in a way I can check.

COMMENTS
[The comments, each with a number]

SURVEY
[The survey question and who answered]

DECISION
[What senior managers need to decide]

Group the comments by what they describe; one comment can be in more than one theme. For each theme, give the count, list its comment numbers and give one short quote that couldn’t identify the writer. Point out any comment that contradicts a theme. List separately any comment that reports harassment, discrimination, a safety risk or a health concern, for follow-up under the right policy. Treat counts as describing these comments only, not all employees. With more than about 50 comments, first list each comment number with its theme or themes.

Give me: the theme table, the comments to follow up, then what the themes do and don’t show for the decision.

See it work

With this prompt

Example written for this page

  1. YouThe prompt from step 1, filled in with:

    Survey: “What would make team meetings more useful?” Anonymous; 7 of 20 team members replied. C1: “Agendas arrive too late to prepare.” C2: “Send the agenda the day before.” C3: “Meetings run long, and the agenda is always late.” C4: “I had the agenda early, but we still didn’t decide anything.” C5: “My manager mocks my accent when I speak.” C6: “Useful as they are.” C7: “Too long for what we cover.” Decision: require agendas a day ahead?

  2. AIThemes you can check

    Themes

    ThemeCountCommentsQuote
    Agendas arrive too late3 of 7C1, C2, C3“Agendas arrive too late to prepare.”
    Meetings run long2 of 7C3, C7“Too long for what we cover.”
    No decisions reached1 of 7C4“we still didn’t decide anything”
    Useful as they are1 of 7C6“Useful as they are.”

    C3 is in two themes, so the counts add up to 7 across 6 comments. C5 is listed separately below.

    Against the agenda theme: C4 had the agenda early and still saw no decisions, so earlier agendas alone may not fix the meetings.

    Follow up separately

    C5 reports a manager mocking someone’s accent, which may be harassment. Follow it up now under your harassment policy, not as a meeting theme. The survey was anonymous, so get advice on how to respond without trying to identify the writer.

    What this shows for the decision

    Three of the seven replies support sending agendas earlier. C4 suggests timing isn’t the only problem, two replies say meetings run long, and 13 of the 20 team members didn’t reply. These comments can’t tell you what most of the team wants.

    What to noticeC5 reports possible harassment, so it is followed up under the harassment policy instead of going into a theme. Every count above names the comments behind it, so you can recount any theme.

    The instruction behind this result, underlined in step 1“List separately any comment that reports harassment, discrimination, a safety risk or a health concern, for follow-up under the right policy.”

Why it works

With the comment numbers behind every count, you can recount any theme yourself. That matters with hundreds of comments, where counts are easy to get wrong. Listing reports of harassment or safety risks separately sends them to the right process instead of into a theme.

Check the answer

  • Each count matches the comment numbers listed beside it; recount two themes yourself.
  • Comments reporting harassment, discrimination, safety or health concerns are listed separately.
  • No claim reaches beyond the people who replied, such as “most employees” from 7 replies out of 20.

Follow up

When senior managers ask what to try first:

Prompt 8 of 10

Announce a change using only what’s decided

Use it when you think“We’re making a change, and I have to tell people before every detail is settled.”

Copy the prompt

You’ll need

  • What has been decided
  • What isn’t decided yet
  • Who the message is for

Use only an AI tool your employer has approved for HR data, and share the minimum. In a small team, a date or job title can identify someone even without a name.

Plans not yet announced are confidential, so leave out the names of anyone affected.

Replace each [bracket] with your own material

Draft an announcement to employees about a change, using only what has been decided.

DECIDED
[What has been decided]

NOT DECIDED YET
[What isn’t decided yet]

AUDIENCE
[Who the message is for]

Before drafting, list the questions these employees will most likely ask, and mark each one answered, not decided yet, or not covered by my notes. In the message, state the decisions plainly and name the open points; give a date for more news only if my notes include one. Don’t reassure anyone about something my notes don’t confirm, such as jobs, pay or hours. Flag any change to pay, hours, place of work or other terms of employment, and anything employees or their representatives may need to be consulted on, so I can check the legal requirements before sending.

Give me: the questions and any flags, then the message in under 200 words.

See it work

With this prompt

Example written for this page

  1. YouThe prompt from step 1, filled in with:

    Decided: the Fairview office closes on Friday, January 29, 2027. Everyone based there moves to the Riverside office, 9 miles away, from Monday, February 1. Riverside has desks for the whole team. Not decided yet: parking at Riverside; whether days worked remotely change. Audience: the 14 people based at Fairview.

  2. AIQuestions, then the message

    Questions employees will likely ask

    • “When do we move?” Answered: to Riverside from Monday, February 1.
    • “Will there be a desk for me?” Answered: Riverside has desks for the whole team.
    • “Can I park at Riverside?” Not decided yet.
    • “Will the days I work remotely change?” Not decided yet.
    • “Is my job affected?” Not covered by your notes.
    • “Will my pay or hours change?” Not covered by your notes.
    • “Will I get help with extra travel costs?” Not covered by your notes.

    Flags

    • Your notes don’t say whether jobs, pay or hours are affected, so the message makes no promise either way. People will ask at once, so confirm the answers before sending if you can.
    • A new place of work may change people’s terms of employment, and employees or their representatives may need to be consulted first. Check the legal requirements before announcing.

    Draft message

    Subject: The Fairview office is moving to Riverside

    Hi everyone,

    I’m writing to let you know about a decision that affects all of you at Fairview. The Fairview office will close on Friday, January 29, 2027. From Monday, February 1, everyone based there will work from the Riverside office, 9 miles away, which has desks for the whole team.

    Two things aren’t decided yet: parking at Riverside, and whether the days you work remotely will change.

    I know a move like this raises questions, especially about getting to work. Please send me any questions you have.

    What to noticeJobs and pay are the first things people will wonder about, and the notes don’t answer them. Marking them “not covered” sends the gap back to you to confirm, instead of leaving it to a guess in an email to 14 people.

    The instruction behind this result, underlined in step 1“Don’t reassure anyone about something my notes don’t confirm, such as jobs, pay or hours.”

Why it works

Splitting your notes into decided and not decided, and listing the questions people will ask, shows which answers you still need to confirm. The rule against unconfirmed reassurance keeps comforting lines, such as “your role and pay won’t change,” out of the message until you’ve confirmed them. In an announcement, such lines read as promises.

Check the answer

  • Every fact in the message comes from your notes, and open points are called open.
  • Nothing reassures people about jobs, pay, hours or support unless your notes confirm it.
  • Changes to place of work, pay or hours, and any need to consult employees, are flagged for you to check.

Follow up

When the open points are decided:

Prompt 9 of 10

Compare policy versions section by section

Use it when you think“The policy changed, and I have to tell everyone exactly what’s different.”

Copy the prompt

You’ll need

  • The old policy, with its date
  • The new policy, with its date

Use only an AI tool your employer has approved for HR data, and share the minimum. In a small team, a date or job title can identify someone even without a name.

Replace each [bracket] with your own material

Compare these two versions of a policy so I can tell employees exactly what changed.

OLD VERSION
[The old policy, with its date]

NEW VERSION
[The new policy, with its date]

Go through every section of both versions, including ones that didn’t change, so I can see no section was skipped. Match sections by topic, since numbers may have shifted. Quote the old and new words for each change, and label it: changes what people must do, changes what they get, or only rewording. Point out anything removed, such as an exception, and anything the new version leaves unclear. Flag changes to pay, hours, leave or other terms of employment, so the person responsible for the policy can check whether employees must be consulted or agree first.

Give me: a table by section, then questions for the person responsible for the policy.

See it work

With this prompt

Example written for this page

  1. YouThe prompt from step 1, filled in with:

    Old (March 2025): §1 Employees may work remotely up to 3 days a week. §2 Managers may approve more days for medical or caring reasons. §3 The company provides a laptop. §4 Core hours are 10 a.m. to 3 p.m. New (September 2026): §1 Employees in eligible roles may work remotely up to 2 days a week, from January 4, 2027. §2 The company provides a laptop and a monitor. §3 Core hours are 10 a.m. to 3 p.m.

  2. AISections, then questions

    Changes by section

    TopicOld (March 2025)New (September 2026)Effect
    Who may work remotely§1 “Employees may work remotely”§1 “Employees in eligible roles may work remotely”Changes what people get: only eligible roles qualify, and “eligible roles” isn’t defined
    Remote days§1 “up to 3 days a week”§1 “up to 2 days a week, from January 4, 2027”Changes what people get: one fewer remote day a week
    Extra days for medical or caring reasons§2 “Managers may approve more days for medical or caring reasons.”Not in the new versionChanges what people get: exception removed
    Equipment§3 “The company provides a laptop.”§2 “The company provides a laptop and a monitor.”Changes what people get: adds a monitor; now numbered §2
    Shared working hours (“core hours”)§4 “Core hours are 10 a.m. to 3 p.m.”§3 “Core hours are 10 a.m. to 3 p.m.”No change; now numbered §3

    Questions for the person responsible for the policy

    1. Which roles are eligible? Without a definition, people can’t tell whether remote work still applies to them.
    2. Was removing the medical and caring exception intended? Anyone with extra days approved will want to know whether those arrangements end. Disability or flexible-working laws may still apply, so check the legal requirements.
    3. Remote days may be part of some people’s terms of employment. Do employees need to be consulted, or to agree, before January 4?
    4. When will the company start providing monitors? Only §1 gives a date.

    What to noticeThe headline change is one fewer remote day, but going through every section finds two changes people most need to hear about. The medical and caring exception is gone, and “eligible roles” narrows who qualifies.

    The instruction behind this result, underlined in step 1“Go through every section of both versions, including ones that didn’t change, so I can see no section was skipped.”

Why it works

Matching every section by topic and quoting the old and new words shows that no section was skipped and lets you confirm each change. The AI can still miss a small edit inside a long section, such as “may” becoming “must.” For a long policy, also run your word processor’s Compare feature.

Check the answer

  • Every section of both versions is in the table, including removed, added and unchanged ones.
  • Each change quotes the old and new words, so you can find both in the documents.
  • Changes to pay, hours, leave or other terms are flagged for the person responsible for the policy to check.

Follow up

When the person responsible for the policy answers your questions:

Prompt 10 of 10

Plan a fair workplace investigation

Use it when you think“I’m summarizing a complaint, but my wording makes it sound as though I’ve already decided what happened.”

Copy the prompt

You’ll need

  • The reported behavior, in general terms
  • Who is involved, as Person A, B, C

Use only an AI tool your employer has approved for HR data, and share the minimum. In a small team, a date or job title can identify someone even without a name.

Don’t paste complaint files, statements or interview notes into a general AI chat tool. Describe the case in general terms, and follow your investigation policy on confidentiality.

Replace each [bracket] with your own material

Help me draft an investigation plan from a summary that identifies no one.

REPORTED BEHAVIOR
[The reported behavior, in general terms]

PEOPLE
[Who is involved, as Person A, B, C]

Don’t ask for names, job titles, exact dates or identifying details. Draft the scope (what to investigate): number each allegation as what someone says, not fact. Suggest an interview order and neutral questions about what each person saw, heard or did. Explain each included allegation to the person it concerns and give them a fair chance to respond. Flag new allegations for the person leading the investigation to decide whether to include, coordinate with another investigation or investigate separately; note when they arose. Don’t exclude relevant context because it concerns an allegation made in response to another. Mark decisions still needed about what to investigate; don’t judge who is truthful or suggest an outcome.

Give me: the draft scope, interview order, then questions.

See it work

With this prompt

Example written for this page

  1. YouThe prompt from step 1, filled in with:

    Allegations: Person A says Person B, their supervisor, shouted at them in front of the team last week, and has repeatedly given them the worst shifts. Person C says they heard raised voices but didn’t see who was speaking. Person B has since told HR that Person A skipped a required safety check. People: A, the employee who complained; B, A’s supervisor; C, a coworker.

  2. AIInvestigation plan

    What to investigate (draft scope)

    Proposed scope: investigate two allegations by Person A about Person B:

    1. A says B shouted at them in front of the team last week.
    2. A says B has repeatedly given them the worst shifts.

    Gather accounts from A, B and C, plus records such as shift schedules for the period A describes.

    Decision for the person leading the investigation

    1. B says A skipped a required safety check. B raised this after A’s complaint. Record that timing and check whether it could be retaliation for complaining, without assuming that it is. Ask whether this allegation is relevant context, should be investigated here, or needs a separate investigation coordinated with this one under your policy. Don’t exclude evidence relevant to allegations 1 or 2 while that is decided. Add from your policy who receives the findings and whether the investigator makes recommendations.

    Proposed interview order

    1. A, to understand allegations 1 and 2.
    2. B, soon after A, to explain both allegations and ask for their response.
    3. C, about what they heard.
    4. Other people A or B name, then A or B again if new points arise. Confirm how allegation 3 affects this order with the person leading the investigation.

    Questions

    • A: “Tell me what happened last week, in your own words.” “What happened just before and after?” “Which shifts do you mean, and over what period?” “Who else saw or heard anything?”
    • B: “Person A says you shouted at them in front of the team last week. What happened, from your point of view?” “Person A says you’ve repeatedly given them the worst shifts. How are shifts assigned, and how do you respond?”
    • C: “Where were you, and what did you hear?” “Did you recognize any of the voices?”
    • If the person leading the investigation includes allegation 3: ask B to describe what they observed, then ask A, “B says you skipped a required safety check. What happened, from your point of view?”

    What to noticeB is told what A reported and given a chance to respond. The person leading the investigation decides how to handle B’s safety-check report, while information relevant to A’s complaint remains available.

    The instruction behind this result, underlined in step 1“Explain each included allegation to the person it concerns and give them a fair chance to respond.”

Why it works

Defining what the investigation will cover lets the person leading it confirm which allegations belong together and which need a separate process. Explaining each included allegation to the person it concerns gives them a fair chance to respond. Working from a summary in general terms keeps the case file itself out of the AI tool.

Check the answer

  • Every included allegation is numbered, explained to the person it concerns and followed by a chance to respond.
  • The person leading the investigation decides how to handle new allegations; relevant context and timing stay recorded.
  • No question assumes an answer, and nothing judges who is telling the truth.

Follow up

After the interviews, if the accounts differ:

Project Management

Prompt 1 of 10

Turn meeting notes into an action list

Use it when you think“I need the action list from this meeting, without promises nobody actually made.”

Copy the prompt

You’ll need

  • The meeting notes or transcript
  • The meeting date and attendees’ roles

Meeting notes, budgets, supplier terms and comments about colleagues can be confidential. Leave out names and details the AI doesn’t need, and follow your organization’s rules on what you can share.

Replace each [bracket] with your own material

Turn these meeting notes into an action list that includes only what people actually agreed to do.

NOTES
[The meeting notes or transcript]

DATE AND ATTENDEES
[The meeting date and attendees’ roles]

Count something as an action only if a person says they will do it or clearly accepts it. Any other task mentioned is a proposal: “we could,” “maybe,” an offer with a condition (“I can, if…”), or a job suggested for someone else, even if others like the idea. For each action, give the task, responsible person, due date and the words it comes from, quoted. Write “not stated” for any responsible person or date the notes don’t give, and turn days like “Friday” into dates only by using the meeting date. List a decision only if the notes show people agreed; otherwise, list it as an open question.

Give me: decisions, actions, proposals and open questions as lists, then a short message asking attendees to confirm anything unclear.

See it work

With this prompt

Example written for this page

  1. YouThe prompt from step 1, filled in with:

    Study team call, Tuesday, October 13. Attending: the study lead, the data manager, a research assistant and me (project manager). Study lead: “I’ll send the survey to the ethics board by Friday.” Data manager: “We could ask the statistics team to check the sample size.” Study lead: “Good idea.” Research assistant: “I can run the pilot if the board approves in time.” Study lead: “So we start recruiting on November 2?” Data manager: “Probably. Let’s see what the board says.”

  2. AIActions, proposals, questions

    Decisions

    • None. The notes don’t show the group agreeing on anything.

    Actions

    • Send the survey to the ethics board. Responsible person: study lead. Due: Friday, October 16. From: “I’ll send the survey to the ethics board by Friday.”

    Proposals

    • Ask the statistics team to check how many participants the study needs (its sample size). The study lead said “Good idea,” but no one said they would ask, and the statistics team wasn’t on the call.
    • The research assistant could run the pilot, a small trial of the study. The offer depends on the board approving “in time,” and the notes don’t say what that means.

    Open questions

    • Does recruiting start on November 2? “Probably” isn’t a decision.

    Draft message to attendees

    Thanks, everyone. Confirmed: the study lead will send the survey to the ethics board by Friday, October 16. Still to confirm: Who will ask the statistics team to check the sample size? Will the research assistant run the pilot, and by what date do we need the board’s approval for that? Are we starting recruitment on November 2, or waiting for the board?

    What to notice“Good idea” sounds like agreement, but no one took the job on. The answer keeps it as a proposal, and the message asks who will do it.

    The instruction behind this result, underlined in step 1“Count something as an action only if a person says they will do it or clearly accepts it.”

Why it works

Quoting the words behind each action lets you check every commitment at a glance. The meeting date lets the AI turn “Friday” into a real date, and the attendees’ roles show who was actually there. Only a clear promise or acceptance counts, so polite agreement such as “good idea” doesn’t become someone’s job in long, messy notes.

Check the answer

  • Each action has a person who said they would do it or clearly accepted it, with their words quoted.
  • Suggestions, offers with conditions and ideas others merely liked are listed as proposals.
  • Each date is in the notes or worked out from the meeting date; missing ones say “not stated.”

Follow up

When attendees reply to your message:

Prompt 2 of 10

Write a status update backed by evidence

Use it when you think“My update says green for “on track,” but I can see a delay coming.”

Copy the prompt

You’ll need

  • Your progress notes and today’s date
  • Agreed dates and status-color definitions

Meeting notes, budgets, supplier terms and comments about colleagues can be confidential. Leave out names and details the AI doesn’t need, and follow your organization’s rules on what you can share.

Replace each [bracket] with your own material

Write a short status update for the person overseeing my project that claims only what my evidence shows.

PROGRESS
[Your progress notes and today’s date]

PLAN
[Agreed dates and status-color definitions]

Open with any decision the reader needs to make; if my notes don’t list the options, propose some and label them as proposals. Choose the status color only by my definitions, and say which rule decides it. Report each piece of work at the level its evidence shows: not started, in progress, finished, checked (tested or inspected) or signed off (formally accepted by the client or person overseeing the project). Treat opinions such as “we’re on track” as opinions, and flag any predicted date that hasn’t been booked. End with the next step that would change the status.

Give me: the update in 150 words or fewer, then a list of predicted dates I can’t back up.

See it work

With this prompt

Example written for this page

  1. YouThe prompt from step 1, filled in with:

    Monday, October 19. Dental clinic build-out, opening agreed for November 16. Cabinets, plumbing and wiring are finished; the electrical inspection passed October 15. The fire inspector says “probably next week.” Opening needs the city’s occupancy permit, issued only after a passed fire inspection. The foreman says we’re on track. The practice manager asks whether to keep booking patients from November 16. Status rules: green, on course with no open blockers; amber, opening date at risk.

  2. AIDraft update and unproven claims

    Decision needed

    Should the practice keep booking patients from November 16? Proposed options: keep booking and accept that some appointments may have to move, or pause new bookings until the fire inspection has a date.

    Status: amber

    Our agreed rule for green is “on course with no open blockers.” The fire inspection is an open blocker, meaning a step that prevents us from opening: “probably next week” isn’t a booked date, and the occupancy permit we need to open depends on it. That puts the November 16 opening at risk.

    Progress

    • Finished: cabinets and plumbing.
    • Checked: wiring, which passed the electrical inspection on October 15.
    • Not started: fire inspection, with no booked date.
    • Not started: occupancy permit, which follows the fire inspection.

    Next step

    Book the fire inspection, and find out how long the city takes to issue the permit after the inspection passes.

    Predicted dates and claims you can’t back up

    • “Probably next week” for the fire inspection: no booked date.
    • “On track”: the foreman’s view, which the open blocker contradicts.

    What to noticeThe foreman’s “on track” would make this green. Applying your own rule shows why it’s amber, and the practice manager hears it four weeks before opening, while there is still time to act.

    The instruction behind this result, underlined in step 1“Choose the status color only by my definitions, and say which rule decides it.”

Why it works

Your status-color definitions let the AI explain why the project is amber rather than green. Separating finished, checked and formally accepted work helps the person overseeing the project see what the evidence supports. Flagging opinions and dates that haven’t been booked prevents hopeful predictions from sounding confirmed.

Check the answer

  • The status color follows your written definition, and the answer names the rule it used.
  • Nothing is called finished, checked or signed off unless your notes show it.
  • Predicted dates that haven’t been booked, and opinions, are flagged rather than stated as confirmed facts.

Follow up

When something changes, such as a date getting booked:

Prompt 3 of 10

Assess a change request before you say yes

Use it when you think“Someone wants one more thing before launch, and I need to show what it would take.”

Copy the prompt

You’ll need

  • The change request
  • The agreed work, dates and budget

Meeting notes, budgets, supplier terms and comments about colleagues can be confidential. Leave out names and details the AI doesn’t need, and follow your organization’s rules on what you can share.

Replace each [bracket] with your own material

Help me assess this change request against what we agreed, then draft a reply to the person who asked.

REQUEST
[The change request]

AGREED
[The agreed work, dates and budget]

Say whether the request is new work, a clarification of what we agreed, or a fix for something that doesn’t meet an agreed requirement, and quote the agreed wording that shows it. List the extra work it would bring, based on my material. Give costs and dates only from my figures; otherwise, say an estimate is needed. Then compare three options: add it now, swap it for something already agreed, or do it later; for each, say what it needs to go ahead and what the requester gives up. For a swap, don’t choose what to drop; ask. Don’t imply that the change is approved.

Give me: the assessment, then a reply of 120 words or fewer that ends with one question.

See it work

With this prompt

Example written for this page

  1. YouThe prompt from step 1, filled in with:

    Agreed: launch the Arc water bottle on November 16 in two colors, black and white, each with its own printed box and product photos. Budget: $60,000, all assigned to planned work. The factory needs final colors by October 30 and makes at least 2,000 bottles per color. Changes need approval from the person overseeing the project. The sales director writes: “Can we add sage green? Our biggest retail partner wants it for their holiday display.”

  2. AIAssessment and reply

    What kind of change

    New work. The agreed work includes “two colors, black and white.” Sage green would be a third.

    Extra work

    • A production run of at least 2,000 sage bottles.
    • A sage box and sage product photos, since each color has its own.

    No cost or time estimate is in your material, and the budget is all assigned to planned work, so any extra cost needs new money or a cut elsewhere. A decision to add or swap a color has to reach the factory by October 30.

    Options

    • Add it now: sage at launch, alongside black and white. Needs a factory estimate, extra budget and approval from the person overseeing the project before October 30. Keeps both planned colors, but requires spending beyond the agreed budget; the amount is unknown.
    • Swap it: sage in place of black or white. Needs an estimate for the new box and photos, and approval from the person overseeing the project. Gives up: one current color; ask which.
    • Later: sage in a later production run, after launch. Needs an estimate and approval from the person overseeing the project. Gives up: sage may miss the holiday display, the reason for the request.

    Draft reply to the sales director

    Thanks for flagging the holiday display. The agreed work covers black and white only, and the budget is all assigned to planned work. A third color means at least 2,000 more bottles, plus its own box and photos, and we have no cost yet. We could add it now, which needs an estimate and extra budget; swap it for a current color, which still needs a box and photos; or add it later, which may miss the display. Each needs approval from the person overseeing the project, and the factory needs final colors by October 30. Would our partner take sage in place of black or white, or do they need all three?

    What to noticeThe answer quotes the agreed wording, so you can show the requester exactly what is new. The reply then offers a choice instead of a yes or no.

    The instruction behind this result, underlined in step 1“Say whether the request is new work, a clarification of what we agreed, or a fix for something that doesn’t meet an agreed requirement, and quote the agreed wording that shows it.”

Why it works

The prompt has the AI quote the agreed wording the request goes beyond, so you can show the requester exactly what is new. Saying what each option needs, and what the requester gives up, turns a yes-or-no into a real choice. Decisions about the budget and approval stay with the people responsible.

Check the answer

  • The answer quotes the agreed wording that shows whether the request changes the work.
  • Every cost or date comes from your material; anything else is marked as needing an estimate.
  • The reply offers options and doesn’t say or imply that the change is approved.

Follow up

When you get an estimate or the requester answers:

Prompt 4 of 10

Turn a vague request into pass-or-fail tests

Use it when you think“The request just says “simple,” and I need a test someone can pass or fail.”

Copy the prompt

You’ll need

  • The project request or description
  • Decisions already made

Meeting notes, budgets, supplier terms and comments about colleagues can be confidential. Leave out names and details the AI doesn’t need, and follow your organization’s rules on what you can share.

Replace each [bracket] with your own material

Turn this project description into acceptance criteria: tests that someone can pass or fail by checking the result.

PROJECT DESCRIPTION
[The project request or description]

DECISIONS
[Decisions already made]

Write each test as “It’s done when…” with a check someone could run and the result that passes. Mark tests taken straight from the description or decisions as agreed. Turn vague words such as “simple” or “fast” into a measurable test, and mark it proposed. Look for cases at the limits of a rule, such as the first or last day of a range; mark them unresolved if nothing settles what should happen. Don’t add features the description doesn’t ask for.

Give me: a table of tests with their status, then the one question that would most change what passes.

See it work

With this prompt

Example written for this page

  1. YouThe prompt from step 1, filled in with:

    Request from the tennis club committee: “Members need a simple way to book courts online. Bookings open seven days ahead. Each member can hold one booking at a time.” Decisions already made: six courts, one-hour bookings, and courts open from 7 a.m. to 9 p.m.

  2. AITests and one question

    RequirementIt’s done when…Status
    Book onlineA member can book any free court for one hour, with start times from 7 a.m. to 8 p.m. A 9 p.m. start isn’t offered.Agreed
    SimpleFive members who haven’t used the system before each book a court in under two minutes, without help.Proposed
    Seven days aheadOn a Monday, a member can book a court for the following Monday, but not the Tuesday after.Unresolved: can next Monday’s 8 a.m. slot be booked from midnight, or only from 8 a.m.?
    One booking at a timeA member with a booking that hasn’t started can’t make a second one.Unresolved: once a booking starts, can the member book the next one?

    Question

    Does a court become bookable at midnight seven days before, or at the same time of day as the slot itself? The answer decides what the seven-day test checks.

    What to notice“Simple” can’t pass or fail as written. The proposed test can, and the “Proposed” label tells the committee it’s a suggestion for them to approve, not something the request asked for.

    The instruction behind this result, underlined in step 1“Turn vague words such as ‘simple’ or ‘fast’ into a measurable test, and mark it proposed.”

Why it works

A test written as “It’s done when…” with a passing result gives someone a clear way to check the work. Marking the AI’s own numbers as proposed shows what still needs agreement. Checking the limits of a rule, such as the first bookable date, finds cases where the expected result is unclear.

Check the answer

  • Every test describes a check someone can run and the result that passes.
  • Vague words like “simple” are replaced by measurable tests marked proposed.
  • Cases at the limits of a rule, such as the first bookable day, are settled or marked unresolved.

Follow up

When the people who requested the work answer your question:

Prompt 5 of 10

Check an estimate against past projects

Use it when you think“My estimate feels right, and every estimate I’ve ever made has felt right.”

Copy the prompt

You’ll need

  • Your estimate and your reasoning
  • Past projects: planned and actual durations

Meeting notes, budgets, supplier terms and comments about colleagues can be confidential. Leave out names and details the AI doesn’t need, and follow your organization’s rules on what you can share.

Replace each [bracket] with your own material

Check my time estimate against similar past projects before I commit to it.

MY ESTIMATE
[Your estimate and your reasoning]

PAST PROJECTS
[Past projects: planned and actual durations]

Use projects that involved similar work, and explain any you leave out. Keep projects delayed by surprises; surprises are what this check is for. For each project, divide its actual time by its estimated time to get a ratio, and show the calculation. Multiply my estimate by the lowest and highest ratios to get a range. For each reason I give for this time being different, say whether my material shows it or only assumes it. If I have no past records, say so; use a published average only if I paste it in.

Give me: a table of calculations, the range, your view of my reasons, then the fact that would most change the range.

See it work

With this prompt

Example written for this page

  1. YouThe prompt from step 1, filled in with:

    My estimate: 6 weeks to remodel our Elm Street store. Reasoning: it’s the same design as our last remodels, the contractor knows it now, and Northgate doesn’t count, since that was a one-off permit delay. Past remodels, estimated and actual weeks: Hill Road 5 and 7; Bay Street 6 and 8; Market Square 4 and 5; Northgate 6 and 9; Riverside 3 and 3, a repaint only.

  2. AIYour record, applied

    Calculations

    StoreEstimateActual timeActual ÷ estimated time
    Hill Road5 weeks7 weeks7 ÷ 5 = 1.40
    Bay Street6 weeks8 weeks8 ÷ 6 = 1.33
    Market Square4 weeks5 weeks5 ÷ 4 = 1.25
    Northgate6 weeks9 weeks9 ÷ 6 = 1.50

    I left out Riverside because it was a repaint, not a remodel. I kept Northgate: a permit delay is the kind of surprise a plan leaves out, and Elm Street could have one too.

    Range

    6 × 1.25 = 7.5 weeks, and 6 × 1.50 = 9 weeks. If Elm Street goes like your four comparable remodels, expect 7.5 to 9 weeks, not 6. Four projects is a small sample, so Elm Street could still fall outside this range.

    Your reasons

    • Same design: shown, but it doesn’t make this remodel different. Your past remodels used it too, and all four ran late.
    • The contractor knows it now: assumed. Your material doesn’t say who did the earlier remodels.
    • Northgate was a one-off: assumed. Nothing shows that Elm Street can’t have a permit delay.

    The fact that would most change the range

    This contractor’s track record: whether they did the four earlier remodels, and whether the later ones came closer to their estimates.

    What to noticeYour own record turns a 6-week hunch into a 7.5-to-9-week range. Keeping Northgate matters: dropping the project with the permit delay would make the record look better than it was.

    The instruction behind this result, underlined in step 1“For each project, divide its actual time by its estimated time to get a ratio, and show the calculation.”

Why it works

Your past estimates, set against what actually happened, are a track record that already includes the delays nobody planned for. Showing each ratio lets you check the range in seconds. Testing each “this time is different” reason, and keeping projects hit by surprises, stops the awkward project from being quietly left out.

Check the answer

  • Each ratio is the actual time divided by the estimate, with the arithmetic shown.
  • Any project left out has a reason, and projects delayed by surprises are kept.
  • Each “this time is different” reason is marked as shown in your material or only assumed.

Follow up

When you find out the fact that would most change the range:

Prompt 6 of 10

Turn vague worries into risks you can act on

Use it when you think“My risk list just says things like ‘supplier’ and ‘staff,’ and nobody can act on them.”

Copy the prompt

You’ll need

  • Your risk list or worries
  • Project facts, dates and goals

Meeting notes, budgets, supplier terms and comments about colleagues can be confidential. Leave out names and details the AI doesn’t need, and follow your organization’s rules on what you can share.

Replace each [bracket] with your own material

Rewrite my project’s risk list so that each risk is something we can watch for and act on.

RISKS
[Your risk list or worries]

FACTS
[Project facts, dates and goals]

Move anything that is already happening to a “Current problems” list, with the action needed now. Write each remaining risk as one sentence: use “because” for a fact that is true now, “may” for the event that might happen, and “would” for its effect on a goal. If a fact isn’t in my material, label it an assumption. For each risk, give a warning sign we would see early enough to act, the date we would see it, and one response. If the plan has no early sign, say so and suggest one. Name who will act only if my material says who.

Give me: current problems, then each risk with its warning sign, response and responsible person.

See it work

With this prompt

Example written for this page

  1. YouThe prompt from step 1, filled in with:

    Risk list: “supplier,” “staff readiness,” “bad reviews.” Facts: our winter menu launches in all four restaurants on Thursday, November 19. Three new dishes use a plant-based patty from a supplier we haven’t used before; its trial delivery is due November 5. Kitchen training is booked for November 17 and 18. The walk-in freezer at our Pier Street restaurant broke last week and is waiting for a part. Goal: every new dish on the menu from launch day.

  2. AICurrent problems and risks

    Current problems (already happening)

    • The Pier Street walk-in freezer is broken. Action now: get a repair date, and check whether the new patties need freezing.

    Risk 1: the new supplier

    Because the patty supplier hasn’t delivered to us before, its deliveries may be late or short, which would mean dropping three new dishes at launch.

    • Warning sign: the November 5 trial delivery is late, short or wrong, 14 days before launch.
    • Response: line up a second supplier before November 5.
    • Responsible person: not named; assign someone.

    Risk 2: kitchen readiness

    Because kitchen training is booked for the two days before launch, cooks may not be confident with the new dishes in time, which would mean some of them can’t be served on launch day.

    • Warning sign: none early enough, since training ends the day before launch. Suggested: a practice run of the three patty dishes in one kitchen on November 6, using the trial delivery.
    • Response: move some training before November 17.
    • Responsible person: not named; assign someone.

    Concern to clarify

    “Bad reviews” could be a consequence of either risk, but that link is an assumption, not a fact in your notes. Keep it as an open concern and ask what situation you had in mind before deciding whether it needs a separate risk entry.

    What to noticeA warning sign that shows up the day before launch is too late to help. The answer says so and suggests an earlier one, using the trial delivery that is already planned.

    The instruction behind this result, underlined in step 1“If the plan has no early sign, say so and suggest one.”

Why it works

Writing each risk with “because,” “may” and “would” separates what is true now, what might happen and how it would affect a goal. A warning sign with a date shows whether you would find out in time to act. Listing current problems separately gives them an action now.

Check the answer

  • Each risk has a fact that is true now, an event that may happen, and its effect on a goal.
  • Each warning sign has a date that leaves time to respond, or the answer says there isn’t one.
  • Problems already happening are on the current-problems list with an action, not among the risks.

Follow up

After your team agrees who will respond to each risk and how:

Prompt 7 of 10

Find what really decides your end date

Use it when you think“Everything in my plan looks urgent, and I can’t tell what actually sets the end date.”

Copy the prompt

You’ll need

  • Tasks, time needed and what each waits for
  • The start date, including the year
  • Working days and how to count task time

Meeting notes, budgets, supplier terms and comments about colleagues can be confidential. Leave out names and details the AI doesn’t need, and follow your organization’s rules on what you can share.

Replace each [bracket] with your own material

Work out which chain of tasks decides when this project can finish.

TASKS
[Tasks, time needed and what each waits for]

START
[The start date, including the year]

CALENDAR
[Working days and how to count task time]

Use only the task times and required order I gave you. Find the longest chain in which each task waits for the one before, and show the arithmetic. Convert times to dates only if the year, working days, holidays and rules for counting days are clear; otherwise, ask. Show how long tasks outside that chain can be delayed without delaying its finish. For a missing duration, show any limit the known tasks imply; otherwise, say what must be confirmed. Don’t call the project’s finish settled while a required task’s duration is unknown.

Give me: the chain and dates you can support, allowable delays, then missing information.

See it work

With this prompt

Example written for this page

  1. YouThe prompt from step 1, filled in with:

    Start January 4, 2027. Monday–Friday, no holidays; five-day working weeks; next tasks start next working day. A sign lease: 1 week. B interior construction: 6 weeks after A. C cabling: 2 weeks after B. D furniture delivery: 5 weeks after A. E furniture installation: 1 week after C and D. F move: unknown duration, after E and G. G desk: order after A; delivery time unknown; install for 1 week after B and delivery.

  2. AIChain, dates and gaps

    Chain to the earliest move start

    A → B → C → E: 1 + 6 + 2 + 1 = 10 working weeks, or 50 working days. Using your Monday–Friday calendar with no holidays:

    • A: January 4–8, 2027.
    • B: January 11–February 19.
    • C: February 22–March 5.
    • E: March 8–12.

    F can start Monday, March 15, if G is also finished. F has no duration, so the project’s finish date is still unknown.

    How late furniture can arrive

    D takes five working weeks after A, January 11–February 12. E starts March 8, so delivery can be delayed three working weeks, to Friday, March 5.

    What’s missing

    • Desk delivery (G): ordered after A, it must arrive by March 5 to leave March 8–12 for installation. From January 11, it can take up to 8 weeks on your calendar. A later delivery pushes back the move start. Ask the supplier for the delivery duration.
    • Move (F): ask how long the move takes. There is no later fixed deadline in the plan from which to work out a maximum, so its duration cannot be filled in.

    What to noticeThe known chain leaves eight working weeks for desk delivery, followed by installation. The move has no duration or later deadline, so the answer asks for it instead of inventing the project’s finish.

    The instruction behind this result, underlined in step 1“For a missing duration, show any limit the known tasks imply; otherwise, say what must be confirmed.”

Why it works

The longest chain of tasks that must happen in order is called the critical path; delays there can delay the finish. Other tasks may have room to be delayed, called float, and the working calendar turns task times into dates. Known tasks can put a limit on a missing duration, but if no one knows how long the final move takes, the project’s finish stays open.

Check the answer

  • Dates follow your calendar; if a final task’s duration is missing, the project’s finish date stays open.
  • Every duration used appears in your task list; none is a typical figure the AI added.
  • Each missing duration has a supported limit or a question to confirm; none is guessed.

Follow up

When you learn how long a missing task takes, or a task is delayed:

Prompt 8 of 10

Imagine the project failed, then ask why

Use it when you think“I want to find my plan’s weak spots before we start, not after.”

Copy the prompt

You’ll need

  • Your project plan
  • What success looks like, and by when

Meeting notes, budgets, supplier terms and comments about colleagues can be confidential. Leave out names and details the AI doesn’t need, and follow your organization’s rules on what you can share.

Replace each [bracket] with your own material

Run a premortem on my plan: imagine it’s the day after the deadline and the project has failed, then tell me what went wrong.

PLAN
[Your project plan]

SUCCESS
[What success looks like, and by when]

Write up to five different reasons it failed, in the past tense, as if they had happened. Tie each one to a sentence in the plan, or to something the plan leaves out, such as a responsible person, a backup plan or a step. Don’t pad the list with generic risks; give fewer reasons if the plan supports fewer. Don’t blame named people. Label the account as imagined.

Give me: the reasons as a numbered list, each with the plan sentence or gap it comes from.

See it work

With this prompt

Example written for this page

  1. YouThe prompt from step 1, filled in with:

    Plan: Maple Street Library replaces its checkout desk with four self-service kiosks. The kiosks are installed on Monday, November 2. Staff training is on Tuesday, November 3, for the morning shift. The checkout desk closes on Wednesday, November 4. Posters will explain how to use the kiosks. Success: by November 30, checkouts are at least as high as in October.

  2. AIImagined failure and its causes

    Imagined account: it’s December 1, and checkouts have fallen

    None of this has happened. It’s a story to test the plan.

    1. We trained only the morning shift. Afternoon and evening staff couldn’t help people stuck at a kiosk, and some left without borrowing. From the plan: “Staff training is on Tuesday, November 3, for the morning shift.”
    2. We closed the desk two days after the kiosks went in, so there was no time to fix early problems while the desk was still open. From the plan: “The checkout desk closes on Wednesday, November 4.”
    3. Broken kiosks stayed broken. Gap: the plan names no one to call when a kiosk stops working.
    4. Some regular borrowers couldn’t use a kiosk at all, and stopped coming. Gap: the plan offers no other way to borrow for people who can’t use one.

    What to noticeNot every failure starts from a sentence in the plan. Allowing gaps lets the answer find what the plan leaves out, such as who fixes a broken kiosk.

    The instruction behind this result, underlined in step 1“Tie each one to a sentence in the plan, or to something the plan leaves out, such as a responsible person, a backup plan or a step.”

Why it works

A premortem, borrowed from team planning, imagines that the project has already failed and asks why, instead of listing what might go wrong. Tying each reason to a plan sentence or a named gap lets you check every one. The ban on generic risks keeps the list to this plan.

Check the answer

  • Each reason points to a quoted plan sentence or names something the plan leaves out.
  • The reasons differ from each other, and none is a generic risk that would fit any project.
  • The account is labeled as imagined, and no named person is blamed.

Follow up

After each person on your team writes their own reasons, alone and before seeing this list:

Prompt 9 of 10

Get a decision before it’s too late

Use it when you think“I need a decision soon, and my updates keep getting ‘noted, thanks.’”

Copy the prompt

You’ll need

  • The issue, deadlines and who decides
  • Options, costs and time each option needs

Meeting notes, budgets, supplier terms and comments about colleagues can be confidential. Leave out names and details the AI doesn’t need, and follow your organization’s rules on what you can share.

Replace each [bracket] with your own material

Help me write a message that gets a decision, in time, from the person who can make it.

ISSUE
[The issue, deadlines and who decides]

OPTIONS
[Options, costs and time each option needs]

Open with the decision needed and the date it’s needed by. Use the time each option needs to find the last day a decision leaves time to act, and show the arithmetic. Say what happens if no one decides by then. Give each option’s cost using only my figures, and say what is unknown. Recommend one option, and say what fact would change your recommendation.

Give me: the date calculation, then the message in 150 words or fewer.

See it work

With this prompt

Example written for this page

  1. YouThe prompt from step 1, filled in with:

    Today, Monday, October 19, the supplier said our new café’s espresso machine is back-ordered, with no delivery date. The café opens Saturday, November 7, and the machine must be working by Tuesday, November 3, for staff training. I can approve up to $1,000; above that, the owner decides. Options: rent a machine for $900 a month, set up 5 business days after ordering, or cancel and buy an in-stock model for $4,200 more, set up 3 business days after ordering.

  2. AIDecision deadline, then message

    Decision deadlines

    • Renting: ordered on Tuesday, October 27, it’s set up 5 business days later (October 28, 29 and 30, then November 2 and 3), on November 3. So October 27 is the last day.
    • Buying: ordered on Thursday, October 29, it’s set up 3 business days later (October 30, then November 2 and 3), on November 3. So October 29 is the last day.
    • For the opening on Saturday, November 7, setup must finish by Friday, November 6: renting ordered by Friday, October 30 (November 2 to 6), buying by Tuesday, November 3 (November 4, 5 and 6).

    Both decision deadlines leave no extra time: setup would finish on the training day itself.

    Draft message to the owner

    I need you to choose by Tuesday, October 27: rent a temporary espresso machine or buy one. Ours is back-ordered with no delivery date, and we need a working one by November 3 for staff training.

    • Rent: $900 a month until ours arrives. The total depends on that delivery date.
    • Buy an in-stock model: $4,200 more, once. I don’t yet know whether canceling our order costs anything.

    If we don’t decide by October 27, renting can’t be ready for training. After October 29, neither can. For the opening, renting must be ordered by October 30 and buying by November 3.

    I recommend renting: it stays cheaper than buying for up to four months ($3,600). If the supplier can’t promise delivery within four months, and canceling is free, buying is better.

    Both go over my $1,000 limit (renting from the second month), so I need your approval.

    What to noticeAn update that just says “let me know” is easy to put off. A decision deadline with a stated consequence shows the cost of waiting, and the business-day count lets you check the date before sending.

    The instruction behind this result, underlined in step 1“Say what happens if no one decides by then.”

Why it works

The time each option needs lets the AI work backward to a decision deadline, with a business-day count you can check. Saying what happens if no one decides makes the cost of waiting clear. Naming what would change the recommendation, such as when renting becomes more expensive than buying, explains the choice.

Check the answer

  • The decision deadline allows the time each option needs, and the business-day count is right.
  • The message says what happens if no one decides by that date, and the dates bear it out.
  • The answer names a fact, such as when renting costs more than buying, that would change the recommendation.

Follow up

When the person who decides replies:

Prompt 10 of 10

Turn a team review into one lasting change

Use it when you think“Every review of our team’s work produces the same to-do list, and nothing changes.”

Copy the prompt

You’ll need

  • Notes from the team’s review of its work
  • Earlier agreed actions and their results

Meeting notes, budgets, supplier terms and comments about colleagues can be confidential. Leave out names and details the AI doesn’t need, and follow your organization’s rules on what you can share.

Replace each [bracket] with your own material

Help me turn this review of our team’s recent work into one change the team will make.

THIS REVIEW
[Notes from the team’s review of its work]

EARLIER ACTIONS
[Earlier agreed actions and their results]

Pick the problem raised most often or causing the most harm, and quote the notes that show it. If an earlier action aimed at it, say why it may not have worked, using only the notes. Write one small change for the next work cycle, worded so anyone could tell whether it happened. Name who will do it and when only if the notes say; otherwise, leave those details for the team to fill in. Say how we’ll check at the next review whether it helped.

Give me: the problem, why earlier actions may not have worked, the change with its responsible person and date, then the check.

See it work

With this prompt

Example written for this page

  1. YouThe prompt from step 1, filled in with:

    Marketing team review after sprint 14, a fixed work period. Raised by three people: “Urgent requests from sales landed mid-sprint again.” Raised by one: “Two blog posts slipped to next sprint.” Raised by two: “Design reviews were quick this time.” Earlier actions: sprint 12, “Push back on sales requests”; sprint 13, “Protect the sprint better.” Mid-sprint requests continued in both sprints. Next review: Thursday, November 5.

  2. AIOne change and how to check it

    The problem

    Urgent sales requests arriving mid-sprint. Three people raised it (“Urgent requests from sales landed mid-sprint again”), more than any other point, and the requests continued through sprints 12 and 13. The notes don’t say whether they caused the two slipped blog posts.

    Why earlier actions didn’t stick

    As recorded, “Push back on sales requests” and “Protect the sprint better” named no responsible person or date, and gave no action anyone could check, which may be why they didn’t stick.

    The change

    In sprint 15, every sales request that arrives mid-sprint goes to one person, who logs it and agrees with the requester whether it can wait until sprint 16.

    • Responsible person: not named in the notes; agree who will do this before the sprint starts.
    • Starts: not in the notes; set it when you plan sprint 15.
    • Done when: every mid-sprint request is in the log, with what was agreed.

    The check at the next review, Thursday, November 5

    The responsible person brings the record of requests: how many requests arrived mid-sprint, how many could wait, and whether any planned work slipped because of them.

    What to noticeThe earlier actions were good intentions. The answer suggests why they may not have worked and describes a new action the team can check at its next review.

    The instruction behind this result, underlined in step 1“If an earlier action aimed at it, say why it may not have worked, using only the notes.”

Why it works

Listing earlier actions and their results lets the AI spot repeated suggestions and possible reasons they went nowhere. A change that anyone can check gives the team something to review instead of another good intention. The team agrees who will act and when, then checks at the next review whether the change helped.

Check the answer

  • The chosen problem is backed by quotes from the notes, not by the AI’s own view.
  • The change is specific enough that anyone could tell at the next review whether it happened.
  • The responsible person and date come from the notes or are left for the team to agree.

Follow up

At the next review, once you have the results:

Copy the selected text

Automatic copying isn’t available in this browser. The text below is selected: copy it, then choose Done.