agentleFS
zirui-songzirui-song/claude-skills/skills/verify-claims/SKILL.md

verify-claims

Skill17 starsChanged 3 months agoMITView on GitHub

Use before the user reads any draft, memo, literature review, or results write-up — runs an adversarial verification pass that hunts factual errors, unsourced claims, and numbers that don't match the underlying code or output files. Every claim is either cited to an exact file:line or marked UNVERIFIED. Trigger phrases include "verify this draft", "check my numbers", "adversarial verification pass", "did you make this up".

13 other projects use a version of this.

Adopted by

Most recently taken up by

The file

We split this skill into the rules it gives an agent

3 rules. 0 are common, 2 are distinctive, and 1 is very specific to this project.

  • 0

    Common

    Most of the 13 edited versions we read keep these.

  • 2

    Distinctive

    This author's own choices. Worth a look.

  • 1

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Very specific:

    Do not fix the prose in this pass.

    line 63
  2. Distinctive:

    Do not mark something VERIFIED because it looks plausible or matches your memory of an earlier run.

    line 65
  3. Distinctive:

    Do not soften an UNVERIFIED into "likely correct." The whole point is the honest label.

    line 67

“Common” is measured: kept by at least 6 in 10 of the 13 edited versions of this file we could read. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Verify Claims

Standing closing step on any written output. Run it before the user reads the draft, not after they catch something.

Inputs (infer if obvious, otherwise ask)

  • Draft file path (.tex, .md, or a section of one), plus any companion documents that quote the same results: the response letter, the online appendix, slides.
  • Where the underlying evidence lives: estimation logs, results/ CSVs, do-files, source PDFs.

Method

Spawn a subagent whose only job is to find errors. Give it the draft and the evidence directories. It must not fix anything, praise anything, or summarize the argument.

For a full paper or several documents, fan out: one subagent per section or document, then merge their lists. A single agent on a whole draft has run out of session before finishing.

For every checkable claim in the draft, it returns one of:

  • VERIFIED — with the exact file:line of the code, log, or output file the number came from, plus the value found there.
  • MISMATCH — draft says X, source says Y. Report both values and the source location.
  • UNVERIFIED — no source located. Say so plainly. Do not reason backward to a source that "probably" produced it.

Claim types to sweep:

  1. Numbers — coefficients, standard errors, t-stats, p-values, sample sizes, F-stats, percentages, dates, sample-window endpoints. Cross-check against the estimation log, not against another sentence in the draft.
  2. Citations — does the cited paper actually make that claim? Check the real paper, not memory. Attribution errors count as factual errors.
  3. Unsourced assertions — institutional facts, "the literature finds…", magnitude claims, anything phrased as established that has no cite and no output file behind it.
  4. Internal consistency — same quantity quoted in abstract, text, and table; table N vs. text N; sign language ("increases") vs. the actual coefficient sign.
  5. Cross-document consistency — the paper, response letter, online appendix and slides quote the same numbers, table/column references and sample; a dropped table leaves no sentence still citing it.
  6. Claimed changes — anything a previous turn said was done (a table regenerated, a number updated, a file edited): open the file on disk and confirm. A "done" that isn't on disk is a MISMATCH.

Output

A single ranked list, most damaging first, before any prose commentary:

| # | Claim (quoted) | Draft location | Status | Source file:line | Found value |

Rank by: would a referee notice, and does it change the paper's conclusion.

Close with a count: n VERIFIED / n MISMATCH / n UNVERIFIED.

Anti-patterns

  • Do not fix the prose in this pass. Show the diff first; the user decides. When fixes are applied afterwards, list every edit as file:line so the user can find each one.
  • Do not mark something VERIFIED because it looks plausible or matches your memory of an earlier run. Open the file.
  • Do not soften an UNVERIFIED into "likely correct." The whole point is the honest label.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.