The AI prompts for research below help with three jobs: mapping a literature you have gathered, critiquing methods, and checking citations. Every prompt works only from material you paste in, because a chat model asked to "find sources" from memory can invent references that look real.
How to use these prompts
Each prompt is ready to copy into ChatGPT, Claude or Gemini. Replace the blanks in double braces, or keep the default after the colon: {{field:public health}} means public health unless you change it.
The rules that matter most:
- Supply the sources. Run your searches in the databases your field uses, then paste abstracts or full texts. The model organizes, compares and critiques. It does not retrieve.
- Allow "I don't know". Anthropic's guide to reducing hallucinations recommends explicitly giving the model permission to admit uncertainty, grounding answers in word-for-word quotes, and having it retract any claim it can't support with a quote. Most prompts below do all three.
- Ask for uncertainty labels. OpenAI's ChatGPT prompting guide suggests this exact line for factual work: "Label anything uncertain with (needs verification)."
- Keep long papers separate. One paper per prompt for extraction, then compare the extractions.
Which prompts help map a literature?
1. Turn a question into search strings
Use it before you open a database. It produces search terms, not sources.
My research question: {{question}}
Field: {{field:public health}}
Databases I will search: {{databases:PubMed and Scopus}}
1. Break the question into its main concepts.
2. For each concept, list synonyms, related terms, spelling variants and
likely controlled vocabulary terms (mark these "check in the thesaurus").
3. Write a Boolean search string for each database, with notes on syntax
I should confirm in that database's help pages.
4. Suggest inclusion and exclusion criteria to decide in advance.
Do not name or cite any papers.
2. Structured extraction from one paper
Run it once per paper, then use the outputs for prompts 3 and 4.
Extract the following from the paper below. Use only the paper. For every
field, include a short word-for-word quote that supports it. If the paper
does not report something, write "not reported". Do not infer.
- Research question or aim
- Design (e.g. RCT, cohort, qualitative interviews)
- Setting and sample (size, who, how recruited)
- Measures or data sources
- Main findings, with numbers where given
- Limitations the authors state
- Limitations the authors do not state (label these "my reading")
Paper:
{{paper_text}}
3. Map papers into themes
Paste the extractions from prompt 2, or the abstracts, with an ID for each.
Below are summaries of {{count:12}} papers on {{topic}}, each with an ID.
{{summaries}}
Group them into themes. For each theme: a one-sentence description, the
paper IDs in it, where they agree, and where they disagree. Then give a
table with one row per paper: ID, design, sample size, main finding, theme.
Use only the summaries above. If a paper fits no theme, say so rather than
forcing it.
4. Find gaps in the set you supplied
Gaps are only gaps relative to what you gathered, and the prompt says so.
Based only on the papers summarized below, identify gaps: populations,
settings, methods or outcomes that none of them cover, and questions where
the findings conflict and no study resolves it.
For each gap, cite the paper IDs that show it. Then note that these gaps are
relative to this set only, and suggest search terms I should run to check
whether other work already fills each one.
{{summaries}}
Which prompts help critique methods?
5. Methods critique of a study
A first pass before journal club or a deeper read.
Critique the methods of this study as a careful peer reviewer in
{{field:psychology}} would.
{{paper_text}}
Cover: whether the design can answer the question, sampling and selection
bias, measurement validity, confounders and how they were handled, the
statistical analysis, missing data, and whether the conclusions go beyond
the results. Quote the passage behind each point. Rank issues by how much
they threaten the main conclusion. Mark anything you're unsure about with
(needs verification).
6. Read a statistical result
For a result section full of numbers you want to interpret correctly.
Explain what this result does and does not tell us, for a reader who knows
basic statistics:
{{result:paste the sentence or table with the numbers}}
Cover: the effect size and whether it is practically meaningful, the
confidence interval, what the p-value means here, and any common
misreadings to avoid. If information needed to interpret it is missing,
say what is missing instead of assuming.
7. Threats to validity in my own design
Use it while you can still change the design.
Here is the design for my planned study:
{{design}}
List threats to internal validity, external validity, construct validity and
statistical conclusion validity. For each threat, say how likely it is in
this design and one practical way to reduce it. End with the two changes
that would strengthen the design most for the least effort.
8. Pre-submission review of my draft
Ask for the hard reviewer before the real one.
Review my manuscript draft as a demanding reviewer for
{{venue:a mid-tier journal in my field}} would.
{{draft}}
Give: a summary of the argument in three sentences (so I can see if it
comes across), major concerns, minor concerns, and the questions a reviewer
would most likely ask. Point to the paragraph for each concern. Do not
suggest new references by name.
Which prompts help check citations?
9. Does the source support the claim?
The most useful citation check. Paste your sentence and the relevant part of the source.
Check whether the source supports each claim I attribute to it.
My text:
{{my_paragraph}}
Source text:
{{source_text}}
For each claim, give: the claim, a word-for-word quote from the source that
supports it, and a verdict (supported, partly supported, not supported,
overstated). If no quote supports a claim, say "No supporting quote found".
Do not use outside knowledge.
10. Find claims that need a citation
For a draft where some sentences carry more weight than their references.
Read this draft and list every sentence that makes a factual or empirical
claim without a citation, or with a citation that seems too general for
the claim. For each, explain what kind of source would support it (e.g. a
systematic review, a primary study, official statistics). Do not suggest
specific papers or authors.
{{draft}}
11. Reference list check
The model can spot formatting and consistency problems. It cannot confirm that a paper exists, so every entry still gets checked by hand.
Check this reference list against {{style:APA 7th edition}}.
{{references}}
For each entry, flag: missing elements (authors, year, title, journal,
volume, pages, DOI), formatting errors, and inconsistencies with the
in-text citations below. Do not add, complete or correct any DOI, page
number or author from memory. Write "verify" next to anything that needs
checking at the source.
In-text citations:
{{in_text}}
12. Compare a summary to the original
For abstracts, press releases or your own notes that may have drifted from what the paper says.
Compare this summary to the original text. List every statement in the
summary that is not supported by the original, is stronger than the
original, or leaves out an important qualification. Quote the original for
each.
Summary:
{{summary}}
Original:
{{original}}
Which prompts help write it up?
13. Plain-language summary
For a lay summary, a grant report or a briefing.
Write a plain-language summary of this paper for {{audience:policy staff with no research background}}.
Under {{words:250}} words. Cover what was studied, how, what was found,
and what it does not show. Use only the paper. No hype words like
"breakthrough" or "proves".
{{paper_text}}
14. Methods section from my notes
Turns notes into a draft you then check against what you actually did.
Draft a methods section from my notes below, following the reporting
conventions for {{design:a cross-sectional survey}}. Use only what is in my
notes. Where a reporting item is missing, insert [MISSING: item] instead of
filling it in.
{{notes}}
How do you verify references the AI touched?
Check every reference yourself, every time. Practical steps:
- Resolve the DOI. Open it and compare title, authors, year and journal. Crossref's REST API returns a DOI's metadata, and Crossref's docs state that "No sign-up is required to use the REST API".
- Check for retractions. Crossref acquired the Retraction Watch database and made it a public resource.
- Read the cited passage. A real paper cited for something it doesn't say is as much a problem as a fake one. Prompt 9 helps, but the final check is your reading.
What do journals say about AI use?
Rules vary by publisher, so read your target journal's policy. The International Committee of Medical Journal Editors (ICMJE recommendations) says journals should require authors to disclose AI-assisted technology use, and that AI should not be listed as an author. The same page warns that "AI can generate authoritative-sounding output that can be incorrect, incomplete, or biased."
If you are doing a systematic review, follow the reporting guideline for it, such as PRISMA 2020, and record any AI assistance in your methods.
Keep your research prompts in one place
Prompts 2 and 9 get run dozens of times in a single review. Saving them with your field and citation style as defaults saves retyping, and the SKILL.md generator turns one into a skill for Claude Code or Codex. For analysis of data you collected, see the data analysis prompts.
Kipt keeps prompts like these as Markdown files in a private GitHub repo you own, with {{field}} and {{style}} left as blanks you fill in, tags per project, and ⌘K search.