AI prompts for researchers
Last updated
For researchers, AI is a reader, critic and translator — never a source. It structures literature you provide, attacks your methodology before reviewers do, disciplines your abstract, drafts review responses, and translates findings for lay audiences. What it cannot do is be trusted for citations: models fabricate plausible references.
Every template below therefore carries the same spine: real material in, structured output back, and explicit markers for anything the material doesn't support.
How to prompt as a researcher
- Never ask for references — ask for structure, critique and synthesis of papers you paste or attach. Verify anything that looks like a citation.
- Name your field's conventions: reporting standards ([APA/CONSORT/PRISMA]), your discipline's methods vocabulary, target journal.
- Use it adversarially: the best research use is pre-review — let it find the weakness before reviewer 2 does.
- Check data policies before pasting unpublished results into consumer tools; use institutional instances where required.
The five templates
A stack of PDFs or abstracts, and you need the comparative matrix your review chapter is built on.
Below are [N] paper abstracts/excerpts I collected on [RESEARCH QUESTION]. Build: 1) a comparison matrix — study, year, design, N, key finding (with effect size where reported), limitation the authors admit; 2) a synthesis in 3 paragraphs: where findings converge, where they conflict (name the studies), what's genuinely open; 3) a list titled "Claims I could not verify from the provided text" for anything you were tempted to add. Rules: only what's in the material. Conflicts are findings, not problems to smooth over. Material: [PASTE ABSTRACTS / EXCERPTS, NUMBERED]
Design is set, data collection looms — the cheapest moment to find the flaw is now.
Act as a methodologically rigorous reviewer in [FIELD]. No politeness padding. Read my design below and deliver: 1) the 3 most serious threats to validity (internal/external/construct — name which), each with the concrete scenario where it bites; 2) which of my assumptions would a hostile reviewer attack first, and the sentence they'd write; 3) the cheapest design change that removes each threat — or the limitation wording if it can't be removed; 4) end with the one question about this study I appear to be avoiding. Design: [PASTE: QUESTION, HYPOTHESES, SAMPLE + RECRUITMENT, MEASURES, ANALYSIS PLAN]
The paper is done; the abstract is 80 words over and somehow still vague.
Rewrite my abstract for [JOURNAL / CONFERENCE], hard limit [N] words, structure: background (1 sentence) → gap (1) → method (2, with N and design) → results (2–3, with the actual numbers: effect sizes, CIs, p where relevant) → implication (1, no overclaim).
Rules: every quantitative claim must appear with its number; delete throat-clearing ("in recent years…"); no "novel", no "to the best of our knowledge". Then list what you cut and why, so I can veto.
My draft:
[PASTE ABSTRACT]
Key numbers, verbatim from the paper:
[PASTE: N, EFFECTS, CIs]
Revisions due. Reviewer 2 was reviewer 2. You need firm, courteous, complete — without the week of dread.
Draft a point-by-point response letter. For each reviewer comment below: 1) restate the point neutrally in one line, 2) our response — agree-and-changed (with manuscript location), partially-agree (what we changed, what not, why), or respectfully-disagree (evidence-based, cordial), 3) the exact manuscript edit, quoted. Tone: grateful without groveling, confident without defensiveness. Where my notes say "disagree", build the strongest evidence-based case from the material I give you — never concede by accident. Comments + my reaction notes: [PASTE: COMMENT 1 + YOUR NOTES, COMMENT 2 + …] Relevant manuscript excerpts: [PASTE]
Funder, press office or repository wants 200 words a non-specialist understands — and your co-authors won't cringe at.
Write a plain-language summary of the finding below. Audience: [INTERESTED NON-SPECIALIST / PATIENT GROUP / POLICY STAFF]. Length: [200] words.
Rules: no jargon without a five-word gloss; the effect stated with its real-world magnitude ("X in 100 people" beats percentages); uncertainty stated plainly ("in this one study of N=…"); no cure/breakthrough vocabulary; nothing the paper doesn't claim.
Then: list the 3 places where simplification cost precision, so I can decide if any matter.
Paper's key claims and numbers:
[PASTE ABSTRACT + KEY RESULTS]
Frequently asked questions
Can I use AI to find papers?
Use dedicated literature tools (Semantic Scholar, Scopus, PubMed, specialized AI search engines with citation grounding) for discovery. General chat models fabricate references fluently — on this page, AI only processes papers you already have.
Is AI-assisted writing allowed in journals?
Most major publishers now allow AI assistance with disclosure and prohibit AI authorship; policies differ in detail. Check your target journal, keep your prompts and drafts (some ask), and disclose per their format. Critique and structuring assistance is broadly uncontroversial.
Can I paste unpublished data into AI tools?
Check your institution's and funder's data policies first — consumer tools may be excluded for unpublished or sensitive data. Institutional/enterprise instances with no-training guarantees are the safer default; the templates work identically there.
Which research tasks does AI do genuinely well?
Adversarial critique of designs and drafts, structural editing under hard constraints, comparative synthesis of provided texts, response-letter drafting, and register translation (expert ↔ lay). It's weakest exactly where trust matters most: sources, novel claims, and field-specific judgment.