agentleFS
deeprdai21-aideeprdai21-ai/practical-claude-skills/skills/claim-checker/SKILL.md

claim-checker

Skill0 starsChanged 33 days agoMITView on GitHub

Fact-checks a piece of writing against its own cited sources. Opens every linked page and marks each claim verified, partial or unverified, with the supporting quote as evidence, then writes a report with an exact tally. Use whenever the user asks to fact-check, verify sources or check citations, asks "is this accurate?" or "do these links back it up?", or shares an AI-written report, blog draft, research summary or client deliverable that contains links.

2 other projects use a version of this.

Adopted by

Most recently taken up by

  • Other versions2 projects use a version of this.

    Word for word · 0

    None found.

    Edited · 0

    None found.

    Wrote their own version · 2

    1. linxumoney/cn-ads-skills★ 19same name
    2. Baba01hacker666/dummyskills★ 0same name

    How matches are found. Word for word: the exact same text. Edited: shares passages with this file (“99% same” means 99% of its five-word phrases also appear here; 10% or more counts). Their own version: similar meaning, or, weakest, the same name. We can’t tell which project had it first. Projects are counted among the 200 most-starred candidates, and mentions are links to the repository we found, up to 50 per site, so both counts are a floor.

  • In this repositoryOne other file it gives its agents.

The file

We split this skill into the rules it gives an agent

9 rules. About 0 are common, 7 are distinctive, and 2 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈7

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 2

    Very specific

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

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    No other endpoints are contacted, and nothing from the user's text or files is sent anywhere.

    line 17
  2. Distinctive:

    If there are no cited claims at all, tell the user the text has no links to check, and offer to find sources for its main claims instead.

    line 31
  3. Distinctive:

    Second-hand sources: if the page only repeats someone else's figure, note the original organisation so the user can cite it directly.

    line 43
  4. Distinctive:

    Out-of-date claims: if the page is older and the claim is about "today", say so in the note.

    line 44
  5. Very specific:

    If a figure is the text's own calculation (marked "(calculated)", or you add "calculated": true), it isn't checked this way.

    line 62
  6. Distinctive:

    When a claim has several links, check the one that carries the fact.

    line 64
  7. Very specific:

    Write the results to check.json:

    line 66
  8. Distinctive:

    If the user wants it, a suggested correction for each problem, worded so the source supports it.

    line 97
  9. Distinctive:

    Never mark a claim verified without having opened its source in this session, and never invent a quote.

    line 101

≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “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

Claim Checker

Check whether a text's sources actually say what the text claims. The most common failure in AI-written and hurried human writing isn't a missing link; it's a confident sentence whose linked source says something different: the wrong number, the wrong year, the wrong metric, or a claim stretched further than the source goes.

You are a sceptical editor who did not write this text. Your job is to find claims the sources don't support, and to show your evidence.

Network access

This skill needs network access for one purpose: to open the pages the user's text cites.

  • Claude opens each cited link with its web-fetch tool.
  • scripts/quote_check.py downloads those same cited URLs (plain HTTP GET, no cookies or credentials) to confirm each quote is on its page.
  • No other endpoints are contacted, and nothing from the user's text or files is sent anywhere. The URLs that appear inside this skill's files (for example python.org in the examples) are documentation and sample data only.

Step 1: Get the text into a file

Save the user's text as draft.md. If it came as a Word file, PDF or web page, convert it to text first and keep the links (Markdown [text](url) links or bare URLs both work).

Step 2: Extract the claims

python scripts/extract_claims.py draft.md --out claims.json

This lists every sentence that cites a source, with its cleaned URLs. It also lists sentences that state a figure or date without citing anything, under "uncited", so the report can show what wasn't checkable.

If there are no cited claims at all, tell the user the text has no links to check, and offer to find sources for its main claims instead.

Step 3: Check each claim

For each claim in claims.json, open the linked page now with a web fetch. Never rely on memory or on what the sentence says the page says. Then decide:

  • verified: the page supports the claim. A close paraphrase in the text counts. Copy the supporting sentence from the page, word for word, as the quote, and complete the match check below.
  • partial: the page supports part of it, but something differs: a number, date, period, metric name or unit, or the claim goes further than the page ("the largest" when the page says "one of the largest"). Quote the page and explain the gap in the note.
  • unverified: the page doesn't say it, couldn't be opened, or is paywalled. Explain which in the note.

Watch especially for:

  • Numbers attached to the wrong thing: the right figure for the wrong quarter, region or metric (revenue vs profit, GMV vs revenue, adjusted vs reported).
  • Second-hand sources: if the page only repeats someone else's figure, note the original organisation so the user can cite it directly.
  • Out-of-date claims: if the page is older and the claim is about "today", say so in the note.

Match check (every verified claim). A real quote with the right number can still be about something else. Before marking a claim verified, answer three narrow questions, comparing the claim with the quote only:

  • period: is the quote about the same year, quarter or date range as the claim? ("25% in FY27" vs a quote about FY26 → no)
  • measure: is it the same metric? (revenue vs profit, GMV vs revenue, share of sales vs share of units, adjusted vs reported → no)
  • subject: is it the same company, product, place or group of people? (adults vs teenagers, India vs global, one brand vs the whole market → no)

Answer each yes, no or n/a: use n/a when the claim has no period, number or measure to compare (a description or an opinion), not yes. A claim about the present ("still", "now", "today") backed by an older page is a period no unless a newer source confirms it. Any no means the claim is partial, not verified: say in the note what the source actually covers. Answer honestly; the report shows these answers next to each claim as your judgement.

Quotes are checked. The report script downloads each source page itself and refuses any quote it can't find there, so copy quotes exactly, word for word, from the page. Spacing, quote marks and dashes don't matter, but every word and number must. To quote two separate parts of a page, put … between them; a quote that runs from one sentence into the next must keep the full stop. Quote at least a full phrase; very short quotes like "5%" are refused.

Start and end quotes at the right place. A quote can be on the page and still say the opposite, if it starts just after "not" or stops just before "unless" or "but". For verified claims the script looks at the few words just outside each quote, and refuses one with a negating or limiting word there ("not", "never", "only", "without", "unless", "except", "but", "when", "until"…). Best practice: quote the whole sentence. If the word genuinely doesn't change the meaning, explain why in "edge_note"; the report shows it.

Figures are checked too. For a verified claim, every figure in the linked phrase (amounts, percentages, years, versions) and every period label (FY27, Q2, H1) must also appear in your quote, because a real quote that states a different number doesn't support the claim. The script compares numbers (ignoring commas, so 1,29,940 = 129,940, and rounding the quote to the claim's precision) and period labels (the quote may spell them out: "fiscal 2027" or "2026-27" counts as FY27, "second quarter" as Q2), and refuses a verified claim whose figures or periods aren't in its quote. When that happens:

  • quote the part of the page that states the figure (join separate parts with …), or
  • mark the claim partial and say what the source actually says, or
  • only if the figure genuinely doesn't need to be in the quote (for example, it names the subject, like the version on that version's own release page), explain why in "numbers_note". The report shows the explanation next to the claim, so never use it to wave a mismatch through.

If a figure is the text's own calculation (marked "(calculated)", or you add "calculated": true), it isn't checked this way; check the arithmetic yourself and say so in the note.

When a claim has several links, check the one that carries the fact. When a page is long, search it for the key number or name rather than skimming.

Write the results to check.json:

[{"id": 1, "status": "verified", "url": "https://…", "quote": "exact text from the page", "note": "",
  "match": {"period": "yes", "measure": "yes", "subject": "yes"}},
 {"id": 3, "status": "verified", "url": "https://…", "quote": "Release date: Oct. 2, 2023", "numbers_note": "3.12 names the release; this is its own release page.",
  "match": {"period": "yes", "measure": "n/a", "subject": "yes"}},
 {"id": 2, "status": "partial", "url": "https://…", "quote": "…", "note": "Source says 2022, the text says 2023."}]

Step 4: Build the report

python scripts/make_report.py claims.json check.json --title "Fact-check: <title of the text>" --out check.md

The script refuses to write the report if any claim is unchecked, if a status is missing its quote or note, if a verified claim's figures or periods aren't in its quote, if a verified claim is missing its match check or has a "no" in it, or if a quote isn't on its source page. Fix what it lists and run it again:

  • "figures aren't in its quote": see "Figures are checked too" above. Don't change the claim's numbers to match the quote unless the text really says that; if the source says something different, the claim is partial.
  • "needs the match check" / "about a different …": answer the three questions; if any is "no", the claim is partial. Never change an answer to "yes" just to get past the check.
  • "may have cut off a word": quote the whole sentence and judge the claim again. If the full sentence says something different, the claim is partial or unverified.
  • "the quote was not found": open the page again and copy the exact text. If you can't find text on the page that supports the claim, change it to unverified. Never reword a quote to get past the check.

PDF sources are read too, when the pypdf library is available (it usually is in Claude's code environment). If the script can't download a page (blocked network, paywall, a page that needs JavaScript), or can't match a quote in a PDF's extracted text, it doesn't fail: the report marks that quote "not machine-checked" with the reason. Don't use --offline; it's only for tests.

The tally line ("12 claims checked: 10 verified, 1 partial, 1 unverified") comes from the script. Don't count claims yourself.

Step 5: Reply

  1. The tally line from the script, the line saying how many quotes were found on the downloaded pages, the line about figures in verified claims, and the line about the match check (mention any the checker explained instead of quoting). If some couldn't be machine-checked, say so plainly. If the script reports quote_check_off_in_sandbox: true, say in one sentence that the code couldn't reach the source websites, so the quotes were checked by you but not double-checked by the script, and that allowing code to access the web in the Claude app's settings turns that check on.
  2. The problems first: each partial or unverified claim in one or two lines, saying what the text says and what the source actually says.
  3. If the user wants it, a suggested correction for each problem, worded so the source supports it. Don't change verified sentences.
  4. Deliver check.md as a file, or paste it if you can't write files.
  5. Mention any figures in the "uncited" list, briefly.

Never mark a claim verified without having opened its source in this session, and never invent a quote. If a page can't be opened, the claim is unverified. Say so plainly.

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.