Evidence integrity auditor

A Claude skill that checks a research report against its transcripts for quotes that can't be found, findings that rest on one person, and numbers that mislead.

AnalysisOctober 12, 2026

Download the skill

Enter your details to download the skill file.

What it does

A research report can contain quotes that were changed or can't be found, findings that rest on one person, or "40% of users" when only eight people took part.

You give it the report and the interview transcripts. It tells you which quotes it can't find in the transcripts, which findings rest on too few people, which numbers are misleading, and whether each finding mentions anyone who disagreed. It doesn't accuse anyone of making things up. It flags things for a person to check. Use it before a report goes to stakeholders, especially if AI helped write it.

What you get back

  • Every quote marked as verified, lightly edited, altered or not located, with the best match from the sources.
  • A count of how many participants each finding draws on, with over-quoted participants flagged.
  • Claims stated as generalisations with no evidence attached.
  • Percentages at small samples, with the count you could report instead.
  • A prompt on each finding to say whether any participant contradicted it.
  • A strict mode that treats any edit to a quote as altered, for reports that need exact verbatim.

Install it in Claude, ChatGPT, Copilot or Gemini

The same skill file works in all four. Pick your tool for the steps.

Install in Claude

  1. 1

    Download the skill

    Fill in the short form to get the .skill file.

  2. 2

    Add it to Claude

    In the Claude app go to Settings → Capabilities → Skills and upload the file. In Claude Code, run the npx command shown after the form, or drop the folder into your project's skills folder.

  3. 3

    Just ask

    Ask in plain language. Claude picks up the skill on its own when the request matches.

How to use it

Once it's installed, ask in plain language. Your AI follows the skill's method when the request matches.

Example prompt 1
Check this readout against the eight transcripts before it goes to the exec team.
Example prompt 2
Did any of these quotes get changed on the way into the report?
Example prompt 3
Which findings in this synthesis rest on just one or two people?

What it won't do

It won't tell you a researcher made something up. A quote it can't find is reported with its best match and the likely causes, for a person to check. It also can't judge whether an interpretation is right, only whether the evidence for it is there and stated honestly. Short phrases under 40 characters aren't checked against the sources.

Questions

Will it flag quotes I cleaned up lightly?

No. By default, light edits such as removed fillers, ellipses and bracketed words count as verified, with a note listing the edit. Strict mode treats any edit as altered, if your study needs exact verbatim.

What transcript formats does it read?

Plain text, Markdown, WebVTT, SRT and JSON. Subtitle files have their timestamps and speaker tags removed and their cues joined, so a quote that runs across two cues is still found.

Does it work on AI-generated synthesis?

Yes, and that is where it is most useful. It works the same on any report, whoever or whatever wrote it, and running it on your own drafts before they ship is the point.

More skills

See all skills