Research and Analysis Prompts

The model cannot tell you what is true. It can help you structure what you are going to check.

This is the pack where getting the premise wrong is most expensive, so it is worth being blunt about what these tools can and cannot do.

In Why Language Models Hallucinate, Adam Tauman Kalai, Ofir Nachum, Santosh Vempala and Edwin Zhang argue that fabrication is not a bug that better models will grow out of — it is a predictable consequence of how models are trained and evaluated. Benchmarks overwhelmingly reward a confident answer and penalise I don't know, so the training process selects for guessing. Their analogy is a student on a hard exam: when unsure, guess, because a blank scores zero and a guess might not.

The consequences are measurable. Dahl, Magesh, Suzgun and Ho, writing in the Journal of Legal Analysis, found hallucination rates of 58% to 88% across leading models on specific legal queries — and, more troublingly, that the models lacked awareness of their own errors and would reinforce a false premise supplied in the question.

The finding that should settle the argument is the follow-up. Stanford's RegLab tested the purpose-built, retrieval-augmented legal research tools sold by LexisNexis and Thomson Reuters — products marketed as having solved this — and found they hallucinate between 17% and 33% of the time. Retrieval helps a great deal. It does not close the gap, and the tools advertised as hallucination-free were not.

Confidence is not a signal

The practical implication is the one people find hardest to internalise: the fluency and confidence of an answer carry no information about whether it is correct. Both are produced by the same process either way. A fabricated citation arrives formatted exactly like a real one, with a plausible author, a plausible journal, and a plausible year.

This matters more for research than for any other use, because research is precisely the case where you are asking about things you do not already know — which means you are least equipped to notice when the answer is wrong.

So every prompt in this pack is built on one rule: work from material you supply, and never generate the facts.

Concretely, that means these prompts will not invent statistics, sources, citations, quotations, or findings. They distinguish what the source said from what you concluded. And where a task appears to require verification, they identify what needs to be checked rather than claiming to have checked it — because the model cannot browse to your source and confirm it, and a prompt that pretends otherwise is worse than useless.

Used that way, the value is real and large. Structuring an enquiry, surfacing what you have missed, organising evidence, and stress-testing your reasoning are all things these tools do genuinely well, and none of them require the model to be an authority on the facts.

Frame the question before you go looking

Break a Big Question Into Researchable Sub-Questions is the first prompt because most research fails at the framing. Should we enter this market is not researchable; the four or five questions underneath it are, and they have different answers with different confidence levels.

It also does the thing that separates research from confirmation: naming, in advance, what finding would change your mind. If nothing would, you are not researching.

Gather, then organise

Synthesize Multiple Sources Into Key Themes is the workhorse. It works across material you supply and — importantly — surfaces where the sources disagree rather than blending them into a smooth consensus that none of them actually support. Disagreement between sources is usually the most informative thing in a pile of reading.

Compare Two Documents and Summarize the Differences is the narrower version: two contract drafts, two versions of a policy, two vendor proposals.

Turn Research Into a Structured Brief produces the readable output, and holds the line that matters — what the evidence supports, what is inference, and what remains unknown, kept visibly separate.

Interrogate what you have

Evaluate a Source's Credibility works through provenance, funding, method, sample, incentive, and whether the claim is primary or a citation of a citation. That last one catches more bad statistics than anything else — a great many widely repeated figures dissolve when you follow them back and find the trail ends at a blog post citing a blog post.

Identify the Claims in a Text and What to Verify is the prompt whose framing explains this whole pack. It cannot fact-check. What it can do is decompose a piece of writing into its constituent factual claims, mark which are load-bearing, and hand you a prioritised list of what to go and confirm. That is a real and useful division of labour: the model does the decomposition, you do the verification.

Stress-Test an Argument or Plan is deliberately built to steelman before it attacks, because a critique of the weakest version of your plan teaches you nothing. It is most valuable run against something you believe.

Do a Root-Cause Analysis pushes past the first plausible explanation, which is usually a symptom with good PR.

Decide, and show your work

Compare Options in a Decision Matrix and Weigh a Decision both carry the same discipline as the scoring prompts elsewhere on the site: they expose which inputs are estimates. A matrix converts judgements into numbers, and a number looks authoritative regardless of what went into it. Making the weights and the guesses visible turns the output into an argument you can examine rather than an answer you have to accept.

Run a SWOT Analysis is included with a caveat it states itself: SWOT is a prompt for thinking, not a conclusion, and it is easy to produce one that is entirely true and completely useless.

Do a Competitive Analysis works strictly from material you gather — it will not tell you a competitor's pricing, headcount, or roadmap, because anything it produced there would be invention.

Analyze Your Stakeholders maps who is affected, who decides, and who can block — the analysis that most often explains why a technically correct recommendation went nowhere.

Build a Well-Reasoned Argument assembles the case, marks where evidence is needed with a placeholder rather than filling it in, and addresses the strongest objection rather than a convenient one.

Research inside the other packs

The same discipline appears wherever a specific domain needs it. Write an Analysis Plan Before You Pull the Data and Estimate a Number You Have No Data For from data analysis. Size Your Market Without Making Up a TAM and Build an Investor Target List and Work Out Who Actually Fits from startup. Find Foundation Funders Worth Applying To and Apply for a Grant or Non-Dilutive Funding, which read public filings rather than mission statements. Research an Account Before You Reach Out from sales, Find Partners Who Already Sell to Your Customers from partnerships, Set Up Social Listening That Tells You Something Useful from social media, and Plan Original Research That Earns Links and Citations from SEO — the one prompt here about producing evidence rather than consuming it.

Where this stops

Everything above, restated: nothing produced with these prompts is verified merely because it is well-organised. Check every factual claim, every figure, and every citation against a primary source before you rely on it or publish it — and treat a confident, well-formatted reference with particular suspicion, because that is exactly what a fabricated one looks like. For anything with legal, medical, financial, or safety consequences, the research is a starting point for a qualified professional, not a substitute for one.

Sources

• Adam Tauman Kalai, Ofir Nachum, Santosh S. Vempala and Edwin Zhang, Why Language Models Hallucinate, September 2025 — training and evaluation reward guessing over expressed uncertainty
• Matthew Dahl, Varun Magesh, Mirac Suzgun and Daniel E. Ho, Large Legal Fictions: Profiling Legal Hallucinations in Large Language Models, Journal of Legal Analysis 16(1), 2024, 64–93 — hallucination rates of 58% to 88% on specific legal queries (Stanford Law School summary)
• Stanford RegLab and HAI, Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools — purpose-built retrieval-augmented legal tools from LexisNexis and Thomson Reuters hallucinate between 17% and 33% of the time