Build a cited research brief

Step 1 — Sharpen the question until it can be answered

"Should we use AI at work?" can't be answered. "Should our five-person agency adopt an AI meeting-notes tool this quarter, and which risks matter?" can. Make the model do the sharpening:

I want to research this question: [your broad question]. Rewrite it as a decidable question — one where evidence could change the answer. Then list the 3–5 sub-questions the answer actually depends on. Ask me anything you need to know about my situation first.

Answer what it asks. Your situation is context the window needs — without it you get the average answer, and you don't want the average, you want yours.

Step 2 — Ground it: search on, receipts required

Now the research prompt — with web search ON. The mechanism this whole lab turns on: without search, the model answers from its weights — a fluent memory of the internet with no date on it. With search, the sources themselves ride into the window, and claims can carry receipts.

Research the question above using web search. Rules: (1) every factual claim gets the source it came from, with a link and the source's date; (2) prefer primary sources — the study, the documentation, the filing — over articles about them; (3) where sources disagree, show me the disagreement instead of picking a winner; (4) anything you believe but cannot source, label OPINION. Give me the findings first, as a list of claim → source pairs. No prose yet.

Step 3 — Build the brief

Now write the brief, one page: THE QUESTION. THE ANSWER, in three sentences. THE CASE, claim by claim, each with its source and date. WHAT THE OTHER SIDE SAYS, sourced the same way. WHAT'S UNCERTAIN. WHAT WOULD CHANGE THE ANSWER. Keep every source attached to its claim — no orphan claims.

It will read wonderfully. You know by now what fluency is — you watched a machine invent a sentence nobody wrote in What AI actually is — so don't stop here.

Step 4 — Audit the citations before you trust them

The signature move. Open a fresh window and paste the brief in:

Audit this brief's citations using web search. For each claim–source pair, open the source and judge: SUPPORTED (it actually says this), STRETCHED (it says something weaker or narrower), or UNSUPPORTED (it doesn't say this, is about something else, or doesn't exist). Return the audit as a table, worst first.

Then the human pass: open the three most load-bearing sources yourself and read the actual sentences. The failure modes this catches — named, because you will meet them: a real link supporting a different claim than the sentence makes; an old number presented as current; a "source" that is itself an unsourced page echoing other AI output. A citation existing is not a citation being right. Checking is the difference between research and confident-sounding fluency.

Fix what the audit caught. Downgrade every STRETCHED claim to what its source actually supports.

Step 5 — Date it and house it

Add one line to the top: "Checked [today's date]." Facts rot — prices, versions, model capabilities, market numbers — and an undated brief rots silently. Then put it where it will live: the project page, the decision doc, wherever the choice gets made. The brief, not the chat, is what you cite in the meeting.

---

Check your understanding: why does a citation that exists still need checking — and why does the audit run in a fresh window?

Use this process for other research. Decidable question → sourced claims → adversarial citation audit → dated brief on a page. That's vendor comparisons, market sizing, the big purchase, the policy question. Fast research was always the easy part; this is how it becomes research you'd act on.