Most prompt libraries are built to get you an answer. The problem at work is rarely getting an answer; it’s knowing whether you can send it. A summary that invents a figure, or an email that promises a date nobody agreed, costs more time than it saved.
So these nine prompts are built the other way round. Each one gets the job done and leaves a trail you can check in seconds: a page reference, a quoted line, a blank where information was missing. Each comes with why it works, how to adapt it, and what a failed answer looks like.
These prompts were written for this article, not benchmarked across models. They use plain instructions rather than any one product’s features, so they paste into ChatGPT, Claude or Gemini. The techniques behind them come from the vendors’ own guidance, cited below. They reduce errors; they do not remove them.
The four moves behind every prompt here
Anthropic’s guide to reducing hallucinations lists a handful of basic techniques, and they turn out to be most of what a good work prompt needs. We use four:
- Permission to say “I don’t know”. The guide says explicitly giving the model permission to admit uncertainty “can drastically reduce false information.”
- Quotes before claims. Asking for word-for-word quotes grounds the answer in the actual text.
- Stay inside the source. Tell the model to use only the material provided, not its general knowledge.
- A trail to check. Have it cite where each claim came from, so you can audit it.
The same guide ends with the caveat this article shares: these techniques reduce hallucinations but “don’t eliminate them entirely.” If you want the reason models fill gaps with confident guesses, here’s why AI makes things up.
Words in square brackets are yours to replace.
1. Summarise a document you can check
Summarise the attached document in [5] bullet points for [a manager who hasn't read it].
After each bullet, give the page or section it comes from.
If anything important is ambiguous or missing, list it separately under "Unclear".
Do not add anything that isn't in the document.Why it works. Page references turn every bullet into a ten-second check. The “Unclear” list gives the model somewhere to put uncertainty instead of smoothing it over.
Adapt it. Change the reader and the length. For a contract, ask for clause numbers instead of pages.
It failed if a bullet has no reference, or the page it names doesn’t say that. Check two bullets at random. If one is wrong, check them all.
2. Answer a question from a policy, and only the policy
Answer the question below using only the text between the lines.
Quote the sentence your answer is based on.
If the text doesn't answer the question, reply "Not covered in the text" and stop.
---
[paste the policy or section]
---
Question: [your question]Why it works. It combines a source restriction, a required quote and an explicit exit. Without the exit, the model will usually find something to say.
Adapt it. Works for handbooks, contracts, terms of service and product specifications. For long documents, paste the relevant section rather than the whole thing, since a long file may be searched in fragments rather than read whole.
It failed if the answer contains a detail that isn’t in the quoted sentence, or it answers a question you know the text doesn’t cover. Test it once with a question you know is out of scope.
3. Turn bullet points into an email without new promises
Write an email to [name / role] using only the points below.
Keep it under [120] words, in a [friendly but direct] tone.
Do not add promises, dates, figures or names that aren't in my points.
After the email, list anything you had to assume.
---
[your points]Why it works. Drafting is where models most readily add plausible filler, and in an email filler can become a commitment. The assumptions list makes any additions visible.
Adapt it. Add “Match the tone of the email I’m replying to:” and paste it below.
It failed if a date, number or offer appears that you didn’t supply, or the assumptions list is empty but the email says something you didn’t.
4. Pull data into a table without guessing
From the text below, extract [company name, contact, contract value, renewal date] for each
[supplier] into a table.
If a value isn't stated, write "—". Do not infer or estimate missing values.
Add a final column quoting the phrase each row came from.
---
[paste text]Why it works. Extraction is where invented values do the most damage, because they look exactly like real ones. The dash rule and the source column make gaps and guesses visible.
Adapt it. Swap the fields. For long lists, ask for the first ten rows, check them, then continue.
It failed if a messy source produces a table with no dashes at all. Real documents are rarely that complete. Check any row whose quoted phrase doesn’t contain the value.
5. Compare two versions for changes in meaning
Compare version A and version B below. List every change in meaning, not wording:
added or removed obligations, changed conditions, numbers, dates or names.
For each change, quote the relevant line from both versions.
If you're unsure whether a change alters the meaning, flag it rather than leaving it out.
---
VERSION A:
[paste]
---
VERSION B:
[paste]Why it works. A word processor’s compare feature finds every edit but can’t tell you which ones matter. This asks the model for the part it is better at, and makes it quote both sides so you can confirm.
Adapt it. For contracts, add “Rank the changes by how much they affect [our side].”
It failed if the list looks short for the size of the edit. Run your word processor’s compare alongside it and check nothing substantive is missing.
6. Get a briefing that separates solid from shaky
Give me a one-page briefing on [topic] for [purpose].
Label each claim [well established], [debated] or [not sure].
For every figure, date or name, say where I could verify it.
Do not cite a source unless you can name it specifically.Why it works. It gives you a checking order. Start with the “not sure” items and the figures, which is where confident errors usually sit.
Adapt it. If your assistant can search the web, add “Link each source you used.” Then open the links.
It failed if everything comes back “well established”, or the sources are vague (“studies show”, “experts say”). Be aware that the labels are the model’s own judgement, not a measurement, so a confident label can still be wrong. Our five-minute fact-checking routine covers what to do next.
7. Turn meeting notes into actions without inventing owners
Turn these meeting notes into a list of actions.
For each: the action, the owner and the deadline, exactly as stated in the notes.
If an owner or deadline wasn't stated, write "not assigned". Do not infer them from context.
Then list any decisions that were discussed but not clearly agreed.
---
[paste notes or transcript]Why it works. Models are good at making notes look tidier than the meeting was. “Not assigned” and the unagreed-decisions list keep the loose ends visible, which is the point of the exercise.
Adapt it. Add “Draft a follow-up message asking for the missing owners,” as a second step.
It failed if every action has a named owner and a date. Real meetings are rarely that tidy.
8. Get a critique instead of a compliment
Review the draft below as [a skeptical client / a new hire / a finance director].
List the five weakest points, most serious first: unclear claims, missing evidence,
gaps in logic, anything this reader would push back on.
Quote the passage each point refers to. Do not rewrite the draft and do not praise it.
---
[paste draft]Why it works. Asked to “review” something, assistants tend to polish and reassure. Naming a skeptical reader and banning rewrites gets you the objections instead.
Adapt it. Ask for three points on a short draft, or ten on a report.
It failed if the points are generic (“could be more concise”) or come without quotes. That usually means it skimmed.
9. Make it ask before it answers
I want [the outcome]. Before you start, ask me up to five questions whose answers would most
change what you produce. Wait for my answers before writing anything.Why it works. Missing context is a common reason output comes back generic. OpenAI’s optimisation guide tells developers to “include relevant context” the model needs; this prompt gets the model to tell you which context that is.
Adapt it. Use it at the start of anything longer than an email: a plan, a proposal, a job ad.
It failed if the questions are ones you’d already answered. That means your request was clear enough, so tell it to proceed.
Keep the ones you use
Pick the two or three that match your week and save them where you’ll find them: a text snippet tool, a note, or your assistant’s project instructions. For the principles behind all of this, including how to give the model the right material, see our practical guide to prompting. And if a long document seems to get worse answers than a short one, the cause is usually the context window, not the prompt.