Prompt Patterns for Research and Analysis You Can Trust

Extraction before synthesis

Never ask for conclusions in the first step. Ask for extraction: pull every claim, figure and date from the source into a table with a quote for each.

You can verify a table quickly. You cannot verify a paragraph of synthesis without redoing the work.

Only after the table is checked should you ask for interpretation, and it should be constrained to the rows in the table.

Structured comparison

Define the criteria yourself, then ask for a matrix scoring each option against each criterion with a justification cell. Undefined criteria produce comparisons that flatter whichever option was described most enthusiastically.

Ask explicitly for the strongest argument against the winning option. The absence of a counter-argument is usually a sign the comparison was decorative.

Then weight the criteria yourself. Weighting is a values judgement and should not be delegated.

Steelman and adversary

Ask for the strongest version of the position you disagree with, then ask for the strongest attack on your own. Both are far more useful than asking whether you are right.

Run the adversarial pass as a separate conversation without your original argument's framing, so the model is not anchored by your preferred conclusion.

Keep the outputs and address them in your final document. This is what turns a memo into a decision paper.

Gap-finding

Ask what a competent reviewer would say is missing, what data would change the conclusion, and what question the document avoids.

Gap-finding is the pattern with the highest hit rate for genuine insight, because it plays to pattern recognition rather than reasoning.

Treat the output as a list of leads to investigate rather than as findings.

Chain of verification

After a draft analysis, ask the model to list every factual claim it made, then independently check each against the supplied source and flag unsupported ones.

This catches a meaningful share of invented specifics, though not all, and it is cheap.

Always spot-check yourself. Self-verification correlates with the original error, so it reduces risk rather than eliminating it.

Knowing when not to use a model

Do not use one where the answer depends on private, current or precise data you have not supplied, where a wrong answer is expensive and unverifiable, or where the reasoning must be defensible to a regulator.

Score the risk before you start: how bad is a confident error, and would you notice it? If the answer is 'bad' and 'no', the model can help you organise the work but must not produce the conclusion.

The grounding pattern in detail

Supply the sources, then instruct: answer only from the material provided, quote the sentence supporting each claim, and list anything the question asks that the sources do not cover. That last clause is what converts a plausible answer into a usable one.

Without it, gaps get filled silently. With it, you get an explicit list of what you still need to look up, which is often the most valuable part of the output.

Keep the sources small enough to check. Twenty pages you can spot-verify beats two hundred you cannot, and verification is the entire point of the exercise.

Adversarial passes catch what summaries miss

Run a second prompt over the first output asking a different question: what is the strongest argument against this conclusion, and what evidence would falsify it? Models are markedly better at criticism than at self-correction.

A third useful pass asks for the assumptions the analysis depends on, listed plainly. Analyses fail at their assumptions far more often than at their arithmetic, and assumptions are easy to review once written down.

Do not run these passes in the same conversation if you can avoid it. A fresh context produces genuinely independent criticism rather than polite agreement with the earlier turn.

Numbers need a separate discipline

Never let the model do arithmetic you have not checked, and never accept a statistic without a source you can open. Ask for the calculation to be shown as a formula with inputs so you can recompute it yourself in seconds.

For anything comparative — growth rates, market sizes, salary ranges — ask for the date and geography of each figure. Undated, unlocated numbers are the most common way a confident analysis becomes wrong.

Where the stakes are real, move the numbers into a spreadsheet and use the model only for the interpretation. That division of labour plays to the strengths of both.

Frequently asked questions

Tools mentioned in this article

More in Prompts & Writing