Most people use AI for research the way they use a blender: put everything in, press the button, hope the result is drinkable. A better prompt does not make the model smarter — it makes the output checkable.
Each prompt below is written to be pasted as-is. Replace anything in [brackets].
1. The claim audit
Use this when a document makes bold statements you need to trust.
Claim audit
Read the document below. List every factual claim that would change my decision if it were wrong. For each claim give: (1) the sentence, (2) what evidence the document itself provides, (3) what evidence is missing, (4) how I could verify it in under 10 minutes. Do not summarise. Do not add claims the document does not make. Document: [PASTE]
2. The disagree-with-me pass
The single most useful prompt I use. It forces the model out of agreement mode.
Steelman the opposite
Here is my position: [POSITION]. Argue the strongest honest case against it in 250 words. Use evidence, not vibes. Then tell me which single piece of evidence would most change your mind, and where I could find it.
3. The three-reading summary
Different readers need different depths. Ask for all three at once and pick.
Layered summary
Summarise the document below at three depths: 1) One sentence a busy executive could repeat from memory. 2) Five bullets for someone who will act on it this week. 3) A 200-word explanation for someone new to the topic, defining any jargon. Flag anything you are unsure about with "[verify]". Document: [PASTE]
4. The numbers-only pass
For reports full of statistics, this stops the prose from hiding the data.
Extract the numbers
From the document below, extract every quantitative claim into a table: metric | value | period | source in document (page/section) | caveat. If a number is an estimate or projection, say so in the caveat column. Do not infer or complete missing values.
5. The meeting transcript cleaner
Two-hour transcript in, decisions and owners out.
Transcript to decisions
From this transcript produce: (1) decisions made, with the exact quote that settles each one; (2) action items with owner and deadline, marked "unassigned" if nobody volunteered; (3) open questions that were raised and not answered; (4) anything said that contradicts an earlier decision. Do not invent owners or deadlines.
6. The pre-mortem
Before you commit to a plan, ask what kills it.
Pre-mortem
It is six months from now and this plan failed: [PLAN]. Write the five most plausible post-mortem findings, ordered by likelihood. For each, name the earliest warning sign I could watch for in the next 30 days, and the cheapest mitigation. Be concrete about failure mechanisms, not generic risks.
7. The citation-first answer
For anything you will publish or forward, make sourcing structural rather than optional.
Citation first
Answer the question below, but structure the answer as a list of claims. Each claim must carry one of: [source: URL], [from provided text], or [unsourced]. If a claim would be [unsourced], either drop it or label it clearly as your own inference. End with a short list of what you could not determine. Question: [QUESTION]
What changed when I switched
Three habits made the biggest difference, and none of them are about “magic words”:
- Ask for the artefact you will use. A table beats prose if you are going to paste it into a sheet.
- Demand the uncertainty. “[verify]” and “[unsourced]” markers turn a confident paragraph into something you can audit in seconds.
- Keep the transcript. When a prompt produces a good result, paste the input/output pair into a note. Prompts that worked once rarely work twice by accident.



