Prompts / Research and analysis
Audit Whether Citations Support Claims
You get a claim-by-claim table showing which statements are backed by the cited evidence and which are overstated or unsupported, with safer rewordings. It works because it makes the AI match each claim to a specific passage instead of trusting that a citation exists.
Worked in our testThe reply followed every step and the format: it quoted each claim, marked "No anchor found" for the survey claim, ran the rule check, gave rewordings and a two-sentence summary. It was loose on verdicts, though. Claim 1 got a double label ("Contradicted and overstated") rather than one of the five options, and "Contradicted" is arguably wrong there, because the authors' caution means the claim is overstated rather than contradicted. The verdict cells also fold the one-sentence reasoning into the label.
The prompt
Act as a careful research integrity reviewer who checks whether cited sources really back up the claims made in a draft.
Task: Audit each claim in {draft_passage} against the supporting evidence supplied in {source_excerpts}, and flag any claim that is unsupported, overstated or contradicted.
Steps:
1. List every factual claim in the draft, quoting it and noting the source it cites.
2. For each claim, find the matching passage in {source_excerpts}. If there is none, mark it "No anchor found" and do not guess from memory.
3. Judge each claim as Supported, Partly supported, Overstated, Not supported or Contradicted, with one sentence of reasoning.
4. Check the draft against {reader_constraints} and flag any breach.
5. Suggest a corrected, more cautious wording for every claim that is not fully supported.
Output: A table with columns Claim, Source, Evidence found, Verdict, Suggested rewording. After the table, give a two-sentence summary of the riskiest problems. Do not invent sources or quotes.
Change the parts in curly braces. Use it before publishing a report, white paper or blog post that leans on research from other people.
What happened when we ran it
We sent
Act as a careful research integrity reviewer who checks whether cited sources really back up the claims made in a draft. Task: Audit each claim in Remote working raises productivity in every industry (Carter, 2023). Our survey of small firms shows staff save an average of 90 minutes a day on commuting. Because of this, companies that adopt hybrid schedules always see profits rise within a year (Okafor, 2024). against the supporting evidence supplied in Carter (2023): In a study of 400 call-centre staff in one UK region, remote workers completed 9% more calls per hour than office-based colleagues. The authors caution that results may not apply to other sectors. Okafor (2024): Among 60 firms that adopted hybrid schedules, 38 reported higher profits after twelve months, 14 reported no change and 8 reported lower profits., and flag any claim that is unsupported, overstated or contradicted. Steps: 1. List every factual claim in the draft, quoting it and noting the source it cites. 2. For each claim, find the matching passage in Carter (2023): In a study of 400 call-centre staff in one UK region, remote workers completed 9% more calls per hour than office-based colleagues. The authors caution that results may not apply to other sectors. Okafor (2024): Among 60 firms that adopted hybrid schedules, 38 reported higher profits after twelve months, 14 reported no change and 8 reported lower profits.. If there is none, mark it "No anchor found" and do not guess from memory. 3. Judge each claim as Supported, Partly supported, Overstated, Not supported or Contradicted, with one sentence of reasoning. 4. Check the draft against Do not make causal claims from correlational data. Do not use words such as always, every or proves. Every statistic must have a cited source. and flag any breach. 5. Suggest a corrected, more cautious wording for every claim that is not fully supported. Output: A table with columns Claim, Source, Evidence found, Verdict, Suggested rewording. After the table, give a two-sentence summary of the riskiest problems. Do not invent sources or quotes.
The model replied
## Claim audit
| Claim | Source | Evidence found | Verdict | Suggested rewording |
|---|---|---|---|---|
| 1. "Remote working raises productivity in every industry" | Carter (2023) | Study of 400 call-centre staff in one UK region: remote workers completed 9% more calls per hour than office-based colleagues. Authors caution results may not apply to other sectors. | **Contradicted** (by the authors' own caveat) and overstated | "One study of 400 call-centre staff in a single UK region found remote workers completed 9% more calls per hour than office-based colleagues (Carter, 2023); the authors caution this may not generalise to other sectors." |
| 2. "Our survey of small firms shows staff save an average of 90 minutes a day on commuting." | None cited (own survey) | No anchor found. Neither Carter nor Okafor mentions commuting time or a small-firm survey. | **Not supported** (within the supplied evidence; no citation or survey details) | "Our survey of small firms found staff saved an average of 90 minutes a day on commuting (n = [X], [survey name/date])." Add sample size, method and a reference or appendix. If the data can't be produced, remove the claim. |
| 3. "Because of this, companies that adopt hybrid schedules always see profits rise within a year" | Okafor (2024) | Of 60 firms adopting hybrid schedules: 38 reported higher profits at 12 months, 14 no change, 8 lower. | **Contradicted** ("always"): 22 of 60 firms did not see a rise. Also the causal link ("Because of this") from commuting time to profits is **not supported** | "In one sample of 60 firms that adopted hybrid schedules, 38 (63%) reported higher profits after twelve months, 14 reported no change and 8 reported lower profits (Okafor, 2024). This is an association and doesn't show hybrid working caused the change." |
## Rule check
- **No causal claims from correlational data:** Breached. "Because of this" links commuting savings to profits with no supporting evidence. Also, "raises productivity" and "adopt... see profits rise" imply causation. Okafor reports outcomes only among adopters, with no comparison group, and Carter's design is not described as randomised in the material supplied, so I can't confirm causal identification.
- **Banned words (always, every, proves):** Breached twice: "every industry" (claim 1) and "always" (claim 3).
- **Every statistic must have a cited source:** Breached. The "90 minutes a day" figure has no citation. Also, the Carter citation is attached to a sweeping claim the source doesn't make.
## Summary
The riskiest problems are the two absolute claims ("every industry", "always"), which are directly contradicted by Carter's own caution about other sectors and by Okafor's data showing 22 of 60 firms had no gain or a loss. The uncited 90-minute statistic and the unsupported "Because of this" causal bridge between commuting time and profits make the argument's logic rest on nothing in the supplied evidence.
Run on Claude Sonnet 5.5, 27 Oct 2026. Automatic check: 4 out of 5.
Mentioned for
Claude Code, Claude
Where it came from
Found on GitHub, shared by Imbad0202: Imbad0202/academic-research-skills.
This is our own version of the idea, written and tested by us. The original may differ.