How to Brief AI for Research Without Getting a Term Paper

Anyone who needs to brief AI for research can skip the term paper in 20 minutes: one question, allowed sources, five fields, UNKNOWN where the record is silent, then stop. Run it in ChatGPT, Claude, or Gemini before you say “write me a full overview.”

A full overview is a seminar you did not enroll in. If you want the loop pattern, read what a skill prompt actually is and keep this chat on the brief.

Brief first. Search second. Paper never, unless you start a new job.

A research brief is not a summary of the internet. It is a decision tool: what you will look for, where, what you will refuse, and when you will stop.

1. Write the question as a question

One sentence with a question mark. “What did this paper claim about sample size, and what did it measure?” Not “tell me about sample sizes in psychology.”

Instruct: “Repeat my question. If I smuggled two questions, list them and stop. Do not merge them. Do not answer yet.”

Merging is how a brief becomes a chapter. Split, or pick one.

If the question is still a pile of notes, turn the pile into a one-page brief and only then come back to research.

2. Name allowed sources, and name the refuse

You list the allowed set. Examples: one uploaded PDF, three URLs you will open, your notes, a date range.

“You may use only these sources. If a claim is not in them, write UNKNOWN. Do not finish the thought from training data. Do not cite a paper I cannot open.”

UNKNOWN is a success state. A fluent fake citation is a failure.

When a PDF is in the set, enforce quote plus page, or it does not count. The brief inherits that law.

3. Force the five-field shape

Now ask for the page. Hard cap.

• Question — the one you already locked.
• Scope — what is in, what is out, in one sentence each.
• Sources — the allowed set, restated, no extras.
• Findings so far — bullets, each tied to a source you named, or UNKNOWN.
• Next fact to fetch — one, not a reading list.

Tell ChatGPT, Claude, or Gemini: “Fill these five. Under 400 words. No history of the field. No ‘scholars have long debated.’ If a finding lacks a source from my list, cut it.”

4. Kill the bibliography that appeared anyway

Models love a reference list. It looks like work.

“If you outputted a bibliography I did not ask for, delete it. Keep only sources I named. If you want a new source, write CANDIDATE plus a URL and wait.”

You open the URL. You add it to the allowed set, or you do not. The model does not sneak it in through a footnote.

Skim Promptcrates news when you need a dated industry fact. Dated pages beat a model’s memory of last year’s launch.

5. One human pass, then park the paper

Read the five fields out loud. If a finding sounds like a lecture, it is padding. Cut it.

Check three traps:

• A scope that quietly widened.
• A finding with no openable source.
• A next-fact list that is actually ten papers.

Fix in your words. Send: “Replace only the lines I changed. Do not expand. Do not start the essay.”

Then stop. If you want a paper later, start a new chat with this brief as law.

Save the question-source-five-field loop so the next topic does not start from “write me a full overview.”

Reusable skill card

Trigger: You are about to ask ChatGPT, Claude, or Gemini to research a topic and you can already see the 1,200-word overview.

Input: One question, an allowed source list, a 400-word cap, and a ban on bibliographies you did not name.

Output: Five fields (question, scope, sources, findings, next fact), UNKNOWN where the set is silent, CANDIDATE URLs parked, no term paper.

Stop: After one human pass. No “shall I write the full review.” No extra sources in footnotes.

Takeaways

• One question. Two questions is two briefs.
• Allowed sources first. Training data is not a source.
• Five fields, under 400 words. A contents page means you failed the cap.
• UNVERIFIED citations get dropped, not trusted.
• Save the briefing loop. Start a new chat if you later want a paper.