TL;DR: Here are 30 AI prompts for academic writing, organized across research questions, argument structure, literature synthesis from sources you paste in, methods, results, discussion, abstracts, and revision. AI chatbots fabricate citations at meaningful, measured rates; verify every reference yourself and disclose AI use according to your journal's or institution's current policy.
What are the best AI prompts for academic writing?
The best AI prompts for academic writing handle structure, not content: sharpening a research question, stress-testing an argument, synthesizing sources you have already pasted in, and tightening prose you have already drafted. None ask AI to generate a finding, invent a citation, or produce a claim you haven't verified yourself.
That constraint is the design principle here. A fabricated fact in a paper doesn't just read badly; it can enter the scholarly record, cited by a reader who trusts the paper, then cited again by someone citing that reader. Every prompt below keeps the model working from text you supply, not what it half-remembers about your field.
The 30 prompts cover eight stages: the research question, argument structure, synthesizing pasted literature, methods, results, discussion, the abstract, and revising against reviewer feedback. This is deliberately not a literature-search or citation-management guide (those are linked below); it's about writing the paper itself, whether that's a thesis chapter, a manuscript, or a course paper.
Is it okay to use AI for academic writing?
Yes, for structure, clarity, and revision. No, for generating the claims, data, or citations that make up the paper's substance. That line holds for a dissertation chapter or a short seminar paper alike, and every prompt below is built around it.
The reason isn't abstract. AI chatbots hallucinate citations at rates that have been measured directly, not guessed at.
The practical rule: never ask AI to generate a citation from memory, and never treat an AI-produced author name, year, or claim as real until you've located and checked the original source. Prompts 9–12 and 30 below build this in directly.
Do you need to disclose AI use in academic writing?
Increasingly, yes; it's now a live norm, not an edge case. The International Committee of Medical Journal Editors (ICMJE) requires authors to disclose AI-assisted writing, to be named in the acknowledgment section (checked at icmje.org, August 2026), and states explicitly that a chatbot cannot be an author, since it can't take responsibility for the work's accuracy.
What structure makes an academic AI prompt trustworthy?
Every prompt below follows the same three-layer structure, built around the failure modes above.
The first layer is a role instruction bounding the model to material you provide: "work only from the text I paste below." Without it, the model blends your input with its training knowledge, the exact condition that invents a statistic or a citation that sounds right but doesn't exist.
The second layer is your actual content: notes, a draft section, a data table, an excerpt you found yourself; the model organizes it into clear structured output but doesn't supply it. The third layer is a specific output format matching your journal's conventions. A vague ask like "improve my methods section" produces vague help; a specific one produces something you can paste into a draft and check.
Neither approach replaces verification. The second is simply much harder to misuse.
| Feature | Open-ended "write my paper" prompt | Source-bounded structured prompt |
|---|---|---|
| Works without pasting your own data, notes, or sources | ||
| Risk of fabricated citations or invented findings | High | Low, if you still verify |
| Output ready to submit without your review | ||
| Matches most journals' and institutions' disclosure expectations |
| Writing stage | Prompts | AI produces | You verify |
|---|---|---|---|
| Research question | 1–4 | Question variants, feasibility notes, significance framing | Whether the question fits your actual data and timeline |
| Argument structure | 5–8 | Thesis stress-test, claim-evidence outlines, transitions | Whether the logic matches what your evidence supports |
| Literature synthesis | 9–12 | Synthesis paragraphs, disagreement flags, from sources you paste | Every claim against the pasted source, every citation independently |
| Methods | 13–16 | Formal methods prose, reproducibility notes, limitations | Accuracy against what you actually did |
| Results | 17–20 | Prose narration, captions, significance-language audit | That no interpretation crept into a results paragraph |
| Discussion | 21–24 | Section skeleton, comparison to prior work, implications | That claims don't outrun what your data shows |
| Abstract | 25–27 | Structured draft, compression, lay summary | That compression didn't drop a hedge or inflate a finding |
| Revision | 28–30 | Line edits, reviewer-response drafts, disclosure statement | Tone, accuracy, and that the disclosure is honest |
How do you sharpen a research question with AI prompts?
A question that's too broad produces a paper that argues everything and proves nothing. These four prompts narrow a topic into something one paper can answer, and check that it fits what you can realistically do.
1. Vague-to-answerable question conversion
Use when you have a topic but not yet a question a paper could actually answer.
Role: Advisor who has watched more papers fail from a vague question than
from weak execution.
Context
- Topic: [what you're interested in]
- Discipline: [sets the conventions]
- Paper type: [thesis chapter, journal article, course paper]
- What you already have: [prior reading, data access, or "starting fresh"]
Task
Turn this into 3 candidate research questions, then recommend one.
Rules
- Each question must be answerable through the writing you're actually
planning to do, not require new data collection you don't have.
- Say plainly if the topic is too broad for the paper type, and give the
narrower version rather than pretending it fits.
- Do not cite authors, studies, or findings. None were given to you, and a
fabricated citation here poisons everything written afterward.
- Flag if the question, as stated, would need a different research design
than what the paper type implies.
Example of the standard I want
Weak: "How does social media affect mental health?"
Strong: "Among the population and timeframe I have data on, does X predict Y,
after accounting for Z?"
Output
3 candidate questions · the recommended one with reasoning · what it
deliberately excludes · what design it implies.
2. Question-versus-topic stress test
Use when you're not sure your "question" is actually a question a reader could answer yes or no, or with a specific finding.
Role: Reviewer whose only job is catching topic statements dressed up as
questions before a committee does.
Context
- Current question: [paste it, exactly as written]
- Discipline: [sets conventions]
- What the paper argues: [one sentence, if you have it]
Task
Test whether this is genuinely a question, and fix it if not.
Rules
- A real question has a specific, falsifiable answer the paper could deliver.
"Explore the role of X" is a topic, not a question; say so directly if
that's what you're given.
- Do not answer the question yourself or suggest what the finding might be.
That is not this prompt's job, and doing it risks planting an assumption.
- If the question is compound (asking two things), say so and offer the split.
- Keep the discipline's usual question conventions (hypothesis-driven vs.
exploratory) rather than imposing a foreign structure.
Example of the standard I want
Weak: "This explores how remote work changed team communication."
Strong: "Does daily message volume per team change after a shift to remote
work, and does that change differ by team size?"
Output
Verdict (question or topic) · the rewritten question if needed · the compound
split if relevant · one line on why the fix works.
3. Question-to-data feasibility check
Use when you need to know whether your question actually fits the data or method you have before you commit pages to it.
Role: Methods-minded advisor who checks feasibility before elegance.
Context
- Research question: [paste it]
- Data or evidence you actually have: [describe it precisely]
- Method or approach: [what you're planning to use]
- Timeline or page limit: [the real constraint]
Task
Check whether this question is answerable with what you described.
Rules
- Work only from what's described. Do not assume data you didn't mention.
- If the question needs something your data can't show (e.g. causal language
from cross-sectional data, or population claims from a narrow sample), say
so specifically and propose the version your evidence actually supports.
- Distinguish "this needs a different method" from "this needs a narrower
question." They call for different fixes.
- Do not invent typical sample sizes or benchmarks for your field.
Example of the standard I want
Weak: "Seems fine, might want more data."
Strong: "Your design is cross-sectional at one timepoint; the question as
worded implies change over time. Either add a second timepoint or
reword the question to an association claim."
Output
Fit assessment · the specific mismatch, if any · the narrower or reworded
version that fits your actual evidence · what a different method would add.
4. Significance and stakes paragraph
Use when you need the "so what" that justifies why the question is worth a reader's time.
Role: Editor who has rejected papers with a technically sound question and no
stated reason anyone should care.
Context
- Research question: [paste it]
- Field/audience: [who reads this]
- What's currently unresolved or underexamined: [in your own words, not a
citation]
- Practical or theoretical stakes: [why the answer would matter]
Task
Draft the significance paragraph for the introduction.
Rules
- Ground the stakes in what you actually know is unresolved, not in a
fabricated "gap in the literature" you haven't verified exists.
- Distinguish theoretical significance (advances understanding) from practical
significance (changes what someone does); name which applies, or both.
- Do not cite a specific study, statistic, or prior finding unless you paste it
in yourself. A generic significance claim you can support beats a specific
one you invented.
- Keep it to what the paper you're actually writing can plausibly deliver.
Example of the standard I want
Weak: "This topic is important and understudied."
Strong: "If X predicts Y as hypothesized, it would give [specific practical
actor] a lever they currently lack for [specific decision]."
Output
The paragraph · one line distinguishing theoretical vs. practical stakes · a
flag if the stakes claim needs a citation you'll need to supply and verify.
How do you structure an argument for an academic paper with AI?
An argument that reads clearly to you can still miss a step for a reader encountering it cold. These four prompts stress-test structure and logic from your own drafted material.
5. Thesis statement stress test
Use when you have a thesis sentence and want to know if it can actually carry a full paper.
Role: Committee member who reads a thesis sentence looking for what it can't
support.
Context
- Thesis statement: [paste it exactly]
- Paper length and type: [thesis, article, course paper]
- Evidence you have: [brief list, not full text]
Task
Stress-test this thesis statement.
Rules
- Check it is arguable (a reasonable reader could disagree), specific (not a
restatement of the question), and scoped to what your evidence list can
actually support.
- Flag any part of the thesis your listed evidence doesn't cover. An
unsupported clause in the thesis is the single most common defense-day
problem.
- Do not suggest new evidence to add. Work only with what you listed.
- If the thesis is really two theses joined by "and," say so.
Example of the standard I want
Weak: "Remote work has changed team dynamics in complex ways."
Strong: "Remote work reduced synchronous communication but increased written
documentation, and this trade-off, not a net loss, explains the
productivity data."
Output
Arguability, specificity, and scope check · unsupported clauses flagged ·
compound-thesis split if relevant · the tightened version.
6. Claim-evidence-warrant section outline
Use when you're drafting a section and want to check every claim actually has support and reasoning behind it, not just a citation dropped nearby.
Role: Logic-focused editor who separates what you claim from what you can
show from what connects the two.
Context
- Section topic: [what this section argues]
- Your claims for this section: [list them, plainly]
- Evidence you have per claim: [your data, or sources you've pasted elsewhere]
Task
Build a claim-evidence-warrant outline for this section.
Rules
- For each claim: state the claim, the evidence that supports it (only what
you gave), and the warrant (why that evidence supports that claim, not just
that it's nearby).
- Flag any claim with no listed evidence, and any evidence with no stated
warrant connecting it to the claim above it.
- Do not add evidence you did not provide, even if you sense it exists.
- Order by logical dependency: if claim 2 needs claim 1 established first,
say so.
Example of the standard I want
Weak: "Claim: X increased. Evidence: [data]."
Strong: "Claim: X increased. Evidence: [data]. Warrant: this data shows X
because [specific reasoning], which rules out [specific alternative]."
Output
The outline, claim by claim · gaps flagged · dependency order · what's missing
before this section is defensible.
7. Strongest counterargument and rebuttal
Use when your argument needs to survive contact with its best objection, not its weakest one.
Role: Devil's advocate who argues the other side as well as its strongest
proponent would, not as a straw man.
Context
- Your argument: [paste your thesis or main claim]
- Your evidence: [brief summary]
- Field: [so the objection is discipline-appropriate]
Task
State the strongest counterargument, then draft your rebuttal.
Rules
- The counterargument must be one a genuine expert would actually raise, not
an easy target. A weak counterargument you demolish is worse than none.
- Do not invent a named scholar, study, or statistic to voice the
counterargument. State it as a position, not a citation you haven't
verified.
- The rebuttal must engage the counterargument's actual substance, not
restate your original claim louder.
- If your evidence genuinely cannot answer the counterargument, say that
directly rather than papering over it.
Example of the standard I want
Weak: "Some might disagree, but the evidence clearly shows X."
Strong: "One could argue Y explains the pattern instead of X. This doesn't
fully hold because [specific reason from your evidence], though it
does mean [specific limitation you should state]."
Output
The strongest counterargument · your rebuttal · any part your evidence
genuinely can't answer, stated honestly.
8. Transition and logical-flow audit
Use when you have drafted sections and need to check the argument actually flows between them, not just within each one.
Role: Reader encountering your paper section by section, checking whether
each one earns the reader's trust for the next.
Context
- Sections in order: [paste the opening and closing paragraphs of each,
labelled]
- Overall argument: [one sentence]
Task
Audit the logical flow between sections.
Rules
- For each junction, state what the previous section established and what
the next section assumes; flag any gap between them.
- Distinguish a genuine logical gap from a merely abrupt transition sentence.
The first needs new content; the second needs a rewritten sentence.
- Do not rewrite the sections' content, only diagnose the junctions and
suggest transition language.
- Flag if a later section silently drops or contradicts something an earlier
section established.
Example of the standard I want
Weak: "Transitions feel a bit choppy in places."
Strong: "Section 3 assumes readers already accept the causal claim from
Section 2, but Section 2 only established correlation. Either soften
Section 3's framing or add the missing step."
Output
Junction-by-junction assessment · gaps flagged as content or transition issues
· suggested transition language for the fixable ones.
How do you synthesize literature from your own pasted sources?
Fabrication risk is highest here, since it's tempting to ask AI to "discuss what the literature says" instead of doing the finding yourself. These four prompts work only from sources you've already located and pasted; none ask the model to search or recall.
9. Synthesis paragraph from pasted excerpts
Use when you have several source excerpts and need them woven into one argumentative paragraph, not a list of summaries.
Role: Synthesist who writes a paragraph that argues something, using sources
as support rather than the paragraph being about the sources.
Context
- Excerpts: [paste 3-5, labelled S1, S2, S3...]
- The point this paragraph needs to make: [your claim]
- Discipline's citation convention: [in-text style, or "leave as [CITE: S1]"]
Task
Synthesize these excerpts into one paragraph supporting the stated point.
Rules
- Work only from the pasted excerpts. Do not add a source, finding, or author
you weren't given, even a well-known one you're confident about.
- The paragraph should read as your argument using sources, not as
"S1 says X. S2 says Y." Integrate them into a single line of reasoning.
- Mark every source reference as [CITE: S1] rather than writing an author-year
citation yourself; you'll replace these with verified, correctly formatted
citations.
- If the excerpts don't actually support the stated point, say so rather than
forcing the synthesis.
Example of the standard I want
Weak: "S1 found X. S2 also found X. S3 disagreed."
Strong: "Findings converge on X in most contexts [CITE: S1][CITE: S2], though
[CITE: S3] suggests this weakens under [specific condition]."
Output
The synthesized paragraph with [CITE: ] placeholders · a note if the excerpts
don't fully support the intended point.
10. Sources-in-service-of-your-argument rewrite
Use when a synthesis draft reads as a summary of other people's work rather than support for your own claim.
Role: Editor who flags "literature-report" paragraphs that describe sources
instead of using them to argue something.
Context
- Draft paragraph: [paste it]
- Your actual claim: [the point this paragraph should be supporting]
- Source excerpts behind the paragraph: [paste them, labelled]
Task
Rewrite this paragraph so the sources serve your claim rather than the other
way around.
Rules
- Do not add any source or finding beyond what's pasted here.
- The claim should open the paragraph; sources should appear as evidence for
it, not as the paragraph's organizing structure.
- Preserve every source's actual finding accurately; don't strengthen a
hedge into a certainty to make it fit your claim better.
- If a source doesn't actually support your claim, flag that rather than
quietly omitting the mismatch.
Example of the standard I want
Weak: "According to S1... Meanwhile S2 argues... S3, on the other hand..."
Strong: "X holds under most conditions studied so far [CITE: S1][CITE: S2],
with one clear exception [CITE: S3] that this paper's design can
actually test."
Output
The rewritten paragraph · which sources support the claim vs. which don't
fully fit.
11. Disagreement detector across pasted sources
Use when your sources seem to conflict and you need to know if that's real or just surface-level.
Role: Synthesist who checks whether a disagreement is genuine before writing
it up as one.
Context
- Sources: [paste findings, labelled]
- What appears to conflict: [your initial impression]
Task
Determine whether these sources genuinely disagree.
Rules
- Check first whether different populations, measures, timeframes, or
definitions explain the apparent conflict; that is not a real disagreement.
- Only the pasted sources. Do not bring in outside knowledge of the debate.
- For a genuine disagreement, state candidate explanations without resolving
it by picking the source you like better.
- If the sources actually agree and you misread them, say that plainly.
Example of the standard I want
Weak: "These sources disagree."
Strong: "S2 and S4 appear to conflict, but S2 measures short-term effects and
S4 measures effects at one year. Not necessarily contradictory."
Output
Genuine vs. apparent disagreement · candidate explanations · how to phrase
this honestly if it goes in your review.
12. Citation-ready prose from a synthesis paragraph
Use when your synthesis reads well but you need it formatted with placeholders ready for your reference manager to fill.
Role: Copyeditor preparing a paragraph for citation insertion, not a citation
generator.
Context
- Synthesis paragraph: [paste it, with [CITE: ] placeholders or source labels]
- Citation style: [APA, MLA, Chicago, discipline-specific]
- Source list: [the sources these placeholders refer to, with full details you
have verified]
Task
Format this paragraph for citation insertion.
Rules
- Do not generate a citation's content (author, year, page) yourself. Insert
placeholders exactly where a citation belongs, keyed to the source you
named, for you or your reference manager to complete.
- Flag any [CITE: ] placeholder in the paragraph that has no matching entry in
your source list; that's a citation you haven't actually sourced yet.
- Preserve the paragraph's argument exactly; this is a formatting pass, not a
content rewrite.
- Note where the style guide requires a citation you don't currently have one
for (e.g. a specific claim needing direct support).
Example of the standard I want
Weak: "(Smith, 2023)" invented from nothing
Strong: "[CITE: S1 — verify against your reference manager entry]"
Output
The paragraph with placeholders · unmatched placeholders flagged · any claim
that needs a citation you don't yet have.
For finding and organizing the sources that feed this stage, see 30 AI prompts for literature review. For keeping citations intact across a large paper set, see how to summarize 50 papers without losing citations.
How do you write a clear methods section with AI prompts?
A methods section has one job: let someone else rerun your study from what you wrote. These four prompts turn your own procedure into formal prose without inventing detail you didn't provide.
13. Procedure notes to formal Methods section
Use when you have working notes on what you did and need publication-register prose.
Role: Methodologist who converts working notes into a reproducible Methods
section without adding anything the notes don't say.
Context
- Your notes: [paste them, however rough]
- Study design: [what type of study this is]
- Target format: [journal's methods conventions, or "standard IMRaD"]
Task
Convert these notes into a formal Methods section.
Rules
- Every sentence must trace to something in your notes. Where your notes are
silent on a standard detail (e.g. exact sample size, tool version), write
[TO SPECIFY] rather than filling it with a plausible default.
- Use precise, past-tense, procedural language; avoid vague verbs like
"analyzed" without saying how.
- Preserve your actual sequence of steps; do not reorder for elegance if it
changes what was actually done first.
- Flag anywhere the notes are ambiguous about what happened.
Example of the standard I want
Weak: "We analyzed the data using standard statistical methods."
Strong: "Data were analyzed using [specific test], with [specific software and
version], testing the null hypothesis that [state it]."
Output
The Methods section · [TO SPECIFY] gaps listed together · ambiguities flagged.
14. Reproducibility audit of a drafted Methods section
Use when you have a Methods draft and want to know if a stranger could actually rerun your study from it.
Role: Independent researcher trying to replicate this study from the Methods
section alone, nothing else.
Context
- Methods section: [paste it]
- Study design: [type]
Task
Audit this Methods section for reproducibility.
Rules
- For each step, state whether a reader could execute it exactly as written,
or what's missing.
- Check standard reproducibility elements: sample recruitment and criteria,
materials/instruments with version or source, procedure sequence, and
analysis specification.
- Do not fill in missing details from field convention. Flag them as missing
instead; assuming standard practice was followed is not your job here.
- Rank gaps by how much they'd block replication versus merely reduce
elegance.
Example of the standard I want
Weak: "Methods seem thorough."
Strong: "Section states 'participants completed a survey' without naming the
instrument. A replicator cannot proceed past this step."
Output
Step-by-step reproducibility check · gaps ranked by severity · what to add.
15. Statistical tests described in plain prose
Use when you need your results section's statistics explained clearly, without the model choosing or inventing which tests apply.
Role: Statistical writer who describes tests that were already run; you do
not choose or recommend which test to use.
Context
- Tests you ran: [name them, with your actual output — coefficients, p-values,
confidence intervals, as reported by your software]
- Design: [what's being compared]
- Audience: [field-standard reporting, or a less technical audience]
Task
Write plain-prose descriptions of these statistical results.
Rules
- Use only the numbers you provided. Do not compute anything, and never
generate a p-value or effect size that wasn't given to you.
- State exactly what the test does and does not establish (e.g. a t-test
result does not establish causation on its own).
- Report effect sizes and confidence intervals alongside significance, not
significance alone.
- If the audience is non-technical, simplify language but keep every
numerical claim identical to what you were given.
Example of the standard I want
Weak: "The difference was statistically significant."
Strong: "Group A scored higher than Group B, t(48) = 2.31, p = .025, d = 0.34,
95% CI [0.05, 0.63]."
Output
Plain-prose description per test · what each test does and doesn't establish
· flags if a number you need wasn't provided.
16. Limitations paragraph grounded in your actual design
Use when you need a limitations section that names real constraints, not boilerplate.
Role: Reviewer writing the limitations section you'd actually want a
reviewer to write, specific to this study.
Context
- Design: [what you actually did]
- Sample: [size, recruitment, who's represented and who isn't]
- Known constraints: [budget, timeline, access, anything that shaped choices]
Task
Draft the limitations paragraph.
Rules
- Ground every limitation in the actual design described, not a generic list
of "limitations every study has."
- State what each limitation prevents you from claiming, specifically, not
just that it exists.
- Distinguish limitations you could have avoided from ones inherent to the
question (e.g. an RCT being impractical or unethical here).
- Do not undersell the study either; a limitations section listing every
conceivable flaw as equally serious is itself misleading.
Example of the standard I want
Weak: "This study has some limitations, including sample size."
Strong: "The convenience sample (n=42, one institution) means findings may not
generalize beyond similar settings; this limits the claim to
association within this population, not a population-wide effect."
Output
The limitations paragraph, each limitation tied to a specific consequence for
the claims the paper can make.
How do you narrate results without overstating them?
Results narration is where hedges quietly disappear and correlation becomes causation. These four prompts describe what your data shows, from output you provide, and check the prose doesn't say more than the numbers do.
17. Results table to prose narration
Use when you have a results table or statistical output and need it described in prose, not interpreted.
Role: Results writer whose job is description, not interpretation. Discussion
of what it means belongs in a different section.
Context
- Results/table: [paste your actual output]
- Research question: [so relevance is clear]
- Reporting format: [your discipline's convention]
Task
Narrate these results in prose.
Rules
- Describe only what the numbers show: direction, magnitude, and statistical
detail exactly as given. Do not explain why the pattern occurred; that
belongs in the discussion section.
- Use hedged language that matches the design: "was associated with," not
"caused," unless your design supports causal language.
- Report null and unexpected results with the same directness as the
hypothesized ones. A dropped null result is one of the most common
distortions in results sections.
- Do not add a number not present in your provided output.
Example of the standard I want
Weak: "The intervention clearly improved outcomes."
Strong: "The intervention group showed a mean improvement of 3.2 points
(SD = 1.1) versus 0.8 (SD = 1.4) in the control group, t(58) = 4.12,
p < .001."
Output
The narration, organized by research question or hypothesis, with every
number traceable to your provided output.
18. Overstatement audit of drafted results prose
Use when you've drafted results prose and want to check it doesn't say more than the underlying numbers support.
Role: Statistically literate reviewer checking language against evidence
strength, sentence by sentence.
Context
- Results prose: [paste your draft]
- Underlying numbers: [the actual statistics behind each claim]
Task
Audit this prose for overstatement.
Rules
- Flag every causal verb ("increases," "causes," "leads to") used where the
underlying design or statistics only support an association claim.
- Flag any claim of certainty ("clearly," "strongly," "conclusively") not
matched by the effect size or confidence interval provided.
- Check that null and marginal results are described as such, not
reinterpreted as "trending toward significance" language your discipline's
norms don't accept.
- Do not rewrite the whole section; flag the specific sentences and give the
corrected wording for those only.
Example of the standard I want
Weak: "This strongly confirms that X drives Y."
Strong: "Flagged: 'strongly confirms' and 'drives' overstate a correlational
finding (r = .21). Suggested: 'X was weakly associated with Y.'"
Output
Flagged sentences with the specific overstatement named · corrected wording ·
sentences that were already appropriately hedged, left alone.
19. Figure and table caption drafting
Use when you need captions that let a reader understand a figure without rereading the whole results section.
Role: Editor who writes captions readers actually use to understand a figure
standing alone.
Context
- Figure/table content: [describe or paste the data it shows]
- What it's meant to demonstrate: [the specific point]
- Style: [journal convention, or standard IMRaD]
Task
Draft the caption.
Rules
- State what the figure shows (data, axes, groups) before stating what it
means; a caption that only interprets without describing fails a reader who
can't read the axes correctly.
- Keep interpretation minimal and hedged consistently with your results
section's own language; a caption should not claim more than the text does.
- Include units, sample size, and error bars/intervals if the figure has them.
- Do not invent data values not present in what you described.
Example of the standard I want
Weak: "Figure 2: Results show clear improvement."
Strong: "Figure 2. Mean symptom scores by group at 12 weeks (n=30 per group,
error bars = 95% CI). Lower scores indicate improvement."
Output
The caption, with description before interpretation, and an alt-text version
for accessibility if your target requires one.
20. Most-important-results selection without cherry-picking
Use when you have more results than your word limit allows and need to choose what leads, honestly.
Role: Editor deciding what a word-limited results section foregrounds, without
hiding inconvenient findings.
Context
- All results: [list everything, including null and unexpected findings]
- Research question: [what the paper is answering]
- Word limit or space constraint: [the real number]
Task
Select and order the results to foreground, within the constraint.
Rules
- Prioritize results that directly answer the stated research question, not
the ones that happen to look most striking.
- A null result directly relevant to the research question must be included,
even briefly, not silently dropped for space.
- If something significant but tangential is cut for space, say so explicitly
rather than pretending it wasn't found.
- State your selection reasoning so it can be checked, not just the final
list.
Example of the standard I want
Weak: "Include the strongest results, cut the rest."
Strong: "Include the primary-hypothesis result (central to RQ) and the
unexpected null on the secondary measure (directly relevant, must be
reported); move the exploratory subgroup finding to supplementary
material with a one-line mention."
Output
Prioritized result list with reasoning · what's cut and why · what must stay
regardless of space.
How do you frame a discussion section with AI prompts?
The discussion is where a paper is most often over-claimed, since it sits furthest from raw numbers. These four prompts apply a standard structure and keep interpretation tied to what the results actually showed.
21. Discussion section skeleton
Use when you're starting the discussion and want the standard structural moves in the right order before you write prose.
Role: Advisor outlining a discussion section's standard moves before drafting
begins.
Context
- Research question: [paste it]
- Key results: [summarize, from your actual results section]
- Field convention: [discipline's typical discussion structure]
Task
Build the discussion section skeleton.
Rules
- Cover the standard moves: restate the main finding plainly, interpret it
against the research question, situate it relative to prior work (only
work you'll paste in separately), state limitations, and state
implications.
- Do not write prior-literature comparisons here; that's a separate prompt
that needs your actual pasted sources.
- Keep each move to a bullet, not full prose yet; this is the skeleton.
- Flag which move is riskiest for overstatement in your specific case (often
the interpretation step).
Example of the standard I want
Weak: "Discuss results, then wrap up."
Strong: "1) Restate: X was associated with Y. 2) Interpret: consistent with
[mechanism], though design can't confirm it. 3) Compare... 4)
Limitations... 5) Implications, scoped to what this design supports."
Output
The skeleton, move by move · the riskiest move flagged · what evidence you
still need to paste in for the comparison step.
22. Compare your findings to prior literature you pasted in
Use when you need to situate your result against specific prior work, using only sources you have.
Role: Synthesist comparing your own finding to prior literature, using only
what's provided.
Context
- Your finding: [state it, with the actual statistics]
- Prior findings: [paste excerpts, labelled]
- Research question: [paste it]
Task
Compare your finding to the prior literature provided.
Rules
- Only the pasted prior findings. Do not bring in other studies from memory,
however well-known you're confident they are.
- State clearly whether your finding replicates, extends, contradicts, or is
simply not comparable to each prior finding (different measure, population,
or design).
- Do not resolve a contradiction with prior work by assuming your study is
right; state candidate explanations instead.
- Keep your own hedging consistent with what your design actually supports.
Example of the standard I want
Weak: "This confirms what prior research has shown."
Strong: "This replicates the direction in [CITE: S2] but with a smaller effect
size, possibly due to [specific design difference]; it contradicts
[CITE: S4], for which no explanation in the current data is
available."
Output
Comparison per prior source · replicate/extend/contradict/not-comparable
label · candidate explanations for contradictions · what remains unresolved.
23. Alternative-explanations paragraph
Use when you need to show you've considered other explanations for your finding before settling on your preferred one.
Role: Skeptical reader generating the alternative explanations a reviewer
would raise, before they raise them.
Context
- Your finding: [state it]
- Your design: [so alternatives are design-appropriate]
- Your preferred explanation: [what you currently argue]
Task
Generate alternative explanations for this finding.
Rules
- Alternatives must be plausible given the actual design (e.g. confounding,
measurement artifact, selection effect, reverse causation where the design
can't rule it out).
- For each, state whether your data can rule it out, partially address it, or
cannot address it at all; the last category needs honest acknowledgment,
not dismissal.
- Do not invent alternatives so weak they're easy to dismiss; that defeats the
purpose.
- Keep your preferred explanation in the mix, evaluated by the same standard
as the alternatives.
Example of the standard I want
Weak: "Other explanations are possible but unlikely."
Strong: "Reverse causation cannot be ruled out by this cross-sectional design;
a longitudinal follow-up would be needed to address it directly."
Output
Alternative explanations, each with whether your data addresses it · which
remain genuinely open · how to phrase that openness honestly.
24. Implications paragraph without overclaiming
Use when you need to state what your finding means for theory or practice, scoped to what it actually supports.
Role: Writer stating implications a reviewer wouldn't need to walk back in
their comments.
Context
- Finding: [state it, with hedging intact]
- Audience: [theoretical, practical, or both]
- Who would act on this: [if practical]
Task
Draft the implications paragraph.
Rules
- Scope every implication to what the design and sample actually support; a
finding in one population is not an implication for a different one without
saying so.
- Separate theoretical implications (what this means for understanding) from
practical ones (what someone should do differently); state which applies.
- Do not recommend a specific action unless your findings actually speak to
that action's effectiveness.
- Flag any implication that would require replication or a different design
before it's actionable.
Example of the standard I want
Weak: "These findings suggest practitioners should adopt this approach."
Strong: "These findings suggest the approach merits testing in [target
setting]; the current sample and design do not yet support a general
recommendation."
Output
The implications paragraph · theoretical vs. practical split · what would
need to be true before a stronger recommendation is warranted.
How do you write an abstract with AI?
An abstract is compression under a hard constraint, and compression is exactly where hedges quietly disappear. These three prompts draft, compress, and translate an abstract from a paper you've already written.
25. Structured abstract from your finished sections
Use when your paper is drafted and you need the abstract, not the other way around.
Role: Abstract writer who compresses a finished paper; you do not add
anything the paper itself doesn't state.
Context
- Introduction/background: [paste it]
- Methods: [paste it]
- Results: [paste it]
- Discussion/conclusion: [paste it]
- Target structure: [structured abstract with headers, or unstructured
paragraph; check your target's requirement]
- Word limit: [the actual number]
Task
Draft the abstract from these sections.
Rules
- Every claim in the abstract must trace to one of the pasted sections.
Nothing new gets introduced here, including a citation, a statistic, or a
stronger version of a hedge.
- Preserve hedging language from the results and discussion exactly; "was
associated with" in the results cannot become "caused" in the abstract.
- Hit the word limit; an abstract 20% over will be cut somewhere by an editor
if not by you.
- Include the specific numbers a reader needs (sample size, key statistic),
not just a qualitative summary.
Example of the standard I want
Weak: "This study examines X and finds important effects."
Strong: "In a sample of 210 adults, X was associated with a 12% reduction in
Y (95% CI [8%, 16%]), consistent with [specific interpretation]."
Output
The abstract at the target word count · a note on anything trimmed for length
that changes emphasis.
26. Word-limit compression without losing hedges
Use when your abstract draft is over the limit and needs cutting without quietly becoming a stronger claim.
Role: Compression editor whose rule is that cutting never increases certainty.
Context
- Draft abstract: [paste it]
- Current word count and target: [both numbers]
- Non-negotiable elements: [what must stay — sample size, key statistic,
specific hedge language]
Task
Compress this abstract to the target word count.
Rules
- Cut words, not certainty. If a cut would turn "was associated with" into
"improved," find a different cut instead.
- Preserve every number and every hedge from the original; compression targets
redundant words and subordinate clauses, not substantive content.
- If the target genuinely cannot be hit without cutting something
substantive, say so and name the trade-off rather than making the cut
silently.
- Keep the non-negotiable elements listed, verified present in the final
version.
Example of the standard I want
Weak: "This study strongly demonstrates X improves Y." (shorter, but wrong)
Strong: "X was linked to improved Y (d=0.3)." (shorter, and still accurate)
Output
The compressed abstract at the target count · confirmation each non-negotiable
element survived · any trade-off you had to flag instead of cutting silently.
27. Plain-language abstract for a non-specialist audience
Use when you need a version a funder's lay board, a policymaker, or a general reader could actually follow.
Role: Science communicator who will not buy clarity with accuracy.
Context
- Technical abstract: [paste your finished version]
- Audience: [policymaker, funder's lay reviewer, general public]
- What they'll use it for: [a decision, general awareness, a report]
Task
Write a plain-language version.
Rules
- Keep every hedge from the technical version, translated into plain words:
"went together with," not "caused," if that's what the original says.
- Keep the population studied explicit; dropping it is the most common
distortion when simplifying for a general audience.
- Define or replace jargon on first use; don't do both for the same term.
- State what the study does not show, as plainly as what it does.
Example of the standard I want
Weak: "Exercise reduces depression." (dropped population and certainty)
Strong: "Among 300 adults already in treatment, those who added supervised
exercise reported slightly fewer symptoms. It doesn't tell us whether
exercise alone would help someone not already in treatment."
Output
The plain-language version · a line noting what precision, if any, had to be
sacrificed for clarity, and why it's acceptable.
How do you use AI for revision and reviewer responses?
The last stage is where AI is genuinely strong: line editing, triaging feedback, and drafting the disclosure your paper needs. These three prompts close the loop.
28. Self-revision line edit pass
Use when your draft is complete and needs a clarity and hedging pass before you submit.
Role: Line editor focused on clarity, redundancy, and hedge accuracy, not
content.
Context
- Draft section: [paste it]
- Known weak spots: [if you already suspect where, name them]
Task
Line-edit this section.
Rules
- Fix redundancy, passive-voice overuse, and unclear referents ("this" with no
clear antecedent). Do not change the substance of any claim.
- Flag, don't silently fix, any place where a hedge seems inconsistent with
what an earlier section established; that needs your judgment, not a quiet
edit.
- Keep field-standard terminology and register; do not simplify jargon that
your target audience expects.
- Return a change list alongside the edited text so you can see exactly what
moved.
Example of the standard I want
Weak: Silently rewriting "may be associated with" to "is linked to."
Strong: "Flagged (not changed): 'may be associated with' in para 2 vs.
'clearly shows' in para 4 for the same finding — reconcile these."
Output
The edited section · a change list · flagged inconsistencies left for you to
resolve.
29. Reviewer comment triage and response draft
Use when you have reviewer comments back and need to organize a response before your resubmission deadline.
Role: Response-letter drafter who separates what changed from what you're
arguing against, clearly, for the editor's benefit.
Context
- Reviewer comments: [paste them, numbered]
- Your planned response to each: [accept and revise, partially address, or
push back, with your reasoning]
- Revised text: [paste, if you've already written it]
Task
Draft the reviewer response letter.
Rules
- Address every comment individually and in order; a merged or skipped
comment is the most common reason editors bounce a resubmission.
- For accepted revisions, quote the new text or describe the specific change
made.
- For pushback, state your reasoning respectfully and specifically; do not
just assert the reviewer is wrong.
- Do not claim a change was made if you haven't actually made it yet in your
draft.
Example of the standard I want
Weak: "We have addressed this concern."
Strong: "We agree and have added a limitations sentence (p. 14, para 2)
acknowledging the sample's geographic constraint."
Output
The response letter, comment by comment · a checklist of which responses still
need the actual manuscript edit made before submission.
30. AI-use disclosure statement for your paper
Use when you used AI anywhere in the writing process and need to disclose it accurately.
Role: Research integrity advisor who writes disclosures specific enough to be
checked, not vague enough to hide anything.
Context
- Where AI was used: [which stages — question framing, outlining, drafting,
line editing, revision]
- Where AI was not used: [e.g. citations, data analysis, findings]
- Tool and approximate dates: [name and version, date range]
- Human verification: [what you personally checked, and how]
- Target policy: [your journal's or institution's current AI policy, if you
have it]
Task
Draft the AI-use disclosure statement.
Rules
- Be specific per stage. "AI assisted with this paper" tells a reader nothing
and is the version most likely to be challenged or deemed insufficient.
- State explicitly that AI was not used for generating findings, data, or
citations, if that's true, since that is the distinction that matters most.
- Name the tool and approximate version; behavior differs meaningfully between
versions.
- If you don't have the target's exact policy in front of you, write to the
most conservative common standard and flag [VERIFY AGAINST TARGET POLICY]
rather than guessing at what's required.
Example of the standard I want
Weak: "AI tools were used to assist in preparing this manuscript."
Strong: "[Tool, version] was used to draft section outlines and perform a
line-edit pass on the final manuscript (March 2026). All citations,
data, and findings were produced independently by the authors and
were not AI-generated. AI output was reviewed and revised by the
authors prior to submission."
Output
The disclosure statement · a per-stage table of use and what wasn't used ·
what to confirm against your target's actual policy before submitting.
Free Chrome Extension
Stop rewriting prompts. Start shipping.
Works with ChatGPT, Claude, Gemini, Grok, Midjourney, Ideogram, Veo3 & Kling. 5.0★ on the Chrome Web Store.
Create An AccountHow do you get the most out of these 30 prompts?
Three practices separate a usable draft from one you'll need to rewrite entirely.
- Paste your own material, every time. The most common failure is skipping the paste and asking AI to work from a bare topic instead, exactly the condition that invents citations and findings from training data rather than your own evidence.
- Verify every citation independently, no exceptions. With 39.8% of AI-generated references erroneous or fabricated in direct testing, treat any AI-produced author, year, or claim as unverified until you've located the source. Use citation software, not AI, for the citations themselves.
- Check disclosure requirements before you submit, not after. Policies differ by journal, funder, and institution, and the field is moving toward more disclosure, not less.
For versioning your prompt use across a whole project, see a reproducible AI research workflow. For a wider tool comparison, the best AI prompt tools for academics and researchers covers how prompt managers fit alongside reference managers.
How Prompt Architects fits this workflow
Every prompt above works in any AI tool you already use. Prompt Architects adds the part that makes a 30-prompt system reusable across a thesis rather than rebuilt every session: a one-click generator that turns a rough instruction into a structured prompt template with the role, context, and rules layers already in place, a prompt library that saves the templates you use, and Contexts, which stores your research question and citation style once so it injects automatically instead of being pasted in each time.
Prompt Architects is free to start, no credit card required. It does not write your paper, verify citations, or replace a reference manager; it makes this guide's structure faster to reuse across a project.
Pick the five prompts that match your current stage, whether that's a stuck research question or a discussion section fighting to stay within your data's actual claims. The discipline that matters most across all 30 isn't the prompt wording; it's pasting your own material in and verifying every citation out.
Start free — save your first academic writing prompt in under two minutes →