Build a Research Brief Generator

Steps

1. 01Pass one: scope the question

   Before any searching, force the model to narrow the question and state what would count as an answer. This is the step everyone skips and it is the one that determines whether the rest is useful.

   Prompt Copy

   I want to research: [your question]. Before answering, do three things:
   1. Restate the question in a narrower, answerable form and explain what you cut.
   2. List the 4-6 sub-questions that together would answer it.
   3. State what kind of evidence would settle each sub-question (statistics, primary documents, expert consensus, first-hand accounts).
   Do not answer the question yet.

2. 02Pass two: gather against the sub-questions

   Take the sub-questions from pass one and research them one at a time, in a tool that can actually search. Doing them one at a time is deliberate: a model asked six questions at once will answer the easy ones well and bluff the hard ones, and you will not be able to tell which is which.

3. 03Pass three: verify before you trust

   Paste the gathered material back and ask for an adversarial read. The instruction to distinguish "the source says this" from "I am inferring this" is the one that catches most fabrication.

   Prompt Copy

   Here is a research draft. For each claim: mark it [SOURCED] if a cited source directly supports it, [INFERRED] if it is your reasoning from the sources, or [UNSUPPORTED] if neither. List every claim you marked UNSUPPORTED at the end. Do not rewrite the draft.
   
   DRAFT:
   [paste]

4. 04Assemble the brief

   Combine the three passes into a fixed format: the narrowed question, findings grouped by sub-question, an explicit "what we still do not know" section, and a confidence rating with reasoning. The unknowns section is the most valuable part and the one a single-prompt approach never produces.

5. 05Test it on something you already know

   Run the whole workflow on a topic you are genuinely expert in. This is the only reliable way to calibrate how much to trust it — you will spot the plausible-sounding errors that you would miss in an unfamiliar field.

If you want to go further

• Add a fourth pass that argues the opposite conclusion and see what survives
• Turn the output format into a document template you reuse

Useful alongside this

Try next