Fact-Checking skill for Claude Code
Trace every claim to a primary source you opened. Claude splits a text into atomic claims, reads laterally to judge each source, and follows every citation trail back to the dataset, paper or transcript it came from. Each claim gets a verdict from accurate to fabricated with a confidence label and a re-checkable link, and AI-generated text gets extra scrutiny for invented citations, quotes and numbers.
- name
fact-checking - size162 lines · ≈2,000 tokens
- files1
- licenseMIT
- updated
What Claude sees before loading it
This description is the only part of the skill that sits in context all the time. Claude reads it to decide when to load the full playbook below.
description: Verifies claims, statistics, quotes, images, and sources with professional fact-checking methods — claim extraction, lateral reading, tracing citation trails to the primary source, checking dates and context, reverse image search, archived versions — and returns a verdict per claim with evidence links and a confidence label. Use when asked to fact-check an article, post, report, or draft, verify a statistic or quote, check whether a source is reliable, debunk or confirm a viral claim, or review AI-generated text for fabricated facts and citations.Try asking
- Fact-check every statistic and quote in this blog post draft before we publish
- Did this senator actually say this? Find the original transcript
- Check whether the citations in this AI-written report are real and say what it claims
Use it when
- A draft, article or report needs its facts verified before publishing
- You need to confirm a statistic, quote or image origin
- AI-generated text may contain fabricated citations or numbers
- A viral claim needs a sourced verdict
Skip it when
- You want a deep critique of one study's methods — use paper-analysis
- You need a survey of research on a topic — use literature-review
How to check the result
Every verdict links to a primary source that someone else can open, and fabricated or unsupported claims list exactly which searches came up empty.
The common mistake it prevents
Treating several outlets that repeat a claim as independent confirmation when they all trace back to one unverified origin.
Done well means: every checkable claim is listed, each verdict rests on the most primary source available (opened and read, not inferred from a search snippet), the evidence links let anyone re-verify, uncertainty is labeled instead of rounded to true or false, and nothing in the check itself is fabricated.
Hard rules
- Never mark a claim verified from memory. Open the source. If you cannot access
sources (no browsing tool), output the claim list with the specific check each needs
and label every verdict
unverified. - Never invent or “reconstruct” a citation, URL, quote, or statistic to support a verdict. A source that cannot be found is evidence of a problem, not a gap to fill.
- Primary beats secondary: the dataset, the study, the transcript, the court filing, the official statement. A news article quoting it is a pointer, not proof.
- Check the claim as stated, including its implied scope (place, time, population, units). A true number used for the wrong year or country makes the claim misleading.
- Report what the evidence shows, not what is plausible.
Workflow
- [ ] 1. Extract checkable claims; prioritize
- [ ] 2. For each claim, identify what would prove or disprove it
- [ ] 3. Evaluate the source making the claim (lateral reading)
- [ ] 4. Trace to the primary source
- [ ] 5. Check context: date, scope, units, quote completeness
- [ ] 6. Assign verdict and confidence; record evidence
- [ ] 7. Write the report1. Extract claims
Split the text into atomic, checkable statements. Skip opinions and predictions, but flag opinions presented as facts.
| Claim type | Example | Primary source to seek |
|---|---|---|
| Statistic | “Unemployment fell to 3.4%” | Statistics office release, table, and reference period |
| Attribution / quote | “X said Y” | Video, transcript, original interview, official record |
| Scientific finding | “Study shows coffee lowers risk” | The paper itself; check design and effect size |
| Event | “The bridge closed in 2019” | Contemporaneous official notices, archived news |
| Legal / policy | “The law requires…” | Statute text, regulation, court opinion |
| Image / video | “Photo from yesterday’s protest” | Earliest instance online; metadata; geolocation |
| Citation | “(Smith et al., 2021)” | DOI or publisher record; the cited passage |
Prioritize by harm and centrality: claims the argument depends on, numbers, health, legal, and financial claims first.
3. Lateral reading (evaluate the source)
Do not judge a site by its own About page or design. Leave it and search what others say:
- Who runs it? Ownership, funding, track record, corrections policy.
- Is the author real and qualified for this specific claim?
- What do independent sources (encyclopedias, established fact-checkers, domain experts) say about the outlet and about this claim?
- Quick checks: domain registration age, whether the “organization” exists outside its own site, whether the content is copied from elsewhere.
4. Trace the citation trail
Follow every “according to” link back until you reach the origin, and read the origin:
- Find the earliest/original publication (search exact phrases in quotes; use date filters; check archives).
- Confirm the origin actually contains the claim, with the same number, scope, and conditions. Telephone-game drift is the most common failure: “up to 40%” becomes “40%”, “in mice” disappears, “associated with” becomes “causes”.
- For papers: confirm the paper exists (DOI resolves, journal record), was not retracted, and that the cited sentence is a finding of the paper rather than a statement from its introduction citing someone else.
- For quotes: find audio/video or the full transcript; check surrounding sentences.
Useful tools and commands:
curl -sLI https://doi.org/10.1000/xyz123 | grep -i '^location' # does the DOI resolve?
curl -s "https://api.crossref.org/works/10.1000/xyz123" | python3 -m json.tool | grep -iE '"title"|"published|update-to'
curl -s "https://archive.org/wayback/available?url=example.com/page×tamp=20230101" # closest archived copy- Wayback Machine and archive.today for deleted or edited pages; compare versions.
- Reverse image search (Google Lens, TinEye, Bing Visual Search, Yandex) for the earliest copy of an image; InVID/WeVerify keyframes for video.
- Official statistics portals for numbers; check the release date and any revisions.
- Established fact-checkers (IFCN signatories) for prior checks; cite but still verify.
5. Context checks
- Time: Is the data current? Was the quote made when claimed? Old images recycled as new are common.
- Scope: country vs. state, all adults vs. a subgroup, nominal vs. inflation-adjusted.
- Denominator: “doubled” from 1 to 2 cases; percent vs. percentage points; absolute vs. relative risk.
- Cherry-picking: does the chosen time window or comparison change the conclusion?
- Quote completeness: ellipses removing qualifiers; question vs. statement.
- Correlation vs. causation: does the source support the causal wording?
6. Verdicts and confidence
| Verdict | Meaning |
|---|---|
| Accurate | Matches the primary source in substance and scope |
| Accurate with caveats | Correct but missing context that a reader needs |
| Misleading | Technically true elements arranged to imply something false |
| Inaccurate | Contradicted by the primary source |
| Unsupported | No reliable source found; cited source does not contain it |
| Fabricated | The cited source, quote, or study does not exist |
| Unverifiable | Cannot be checked with available sources (state what would settle it) |
Confidence: High (primary source read, independent corroboration), Medium (primary source read, no corroboration; or strong secondary sources only), Low (indirect evidence, conflicting sources, or limited access).
Checking AI-generated text
LLM output fails in characteristic ways; check these first:
- Citations: resolve every DOI and URL; search titles in Crossref/Google Scholar. Plausible author-year-journal combinations that do not exist are common.
- Exact numbers and dates, especially round or highly specific ones without a source.
- Quotes attributed to real people; search the exact string.
- Legal citations (case names, statute sections); check an official court or legislative database.
- Superlatives and firsts (“the first”, “the largest”).
Report template
# Fact-check: <title of piece> (checked 2025-10-14)
Summary: 12 claims checked — 7 accurate, 2 accurate with caveats, 1 misleading, 1 inaccurate, 1 fabricated citation.
| # | Claim (verbatim, with location) | Verdict | Confidence | Evidence (primary source + link) | Note / suggested fix |
|---|---------------------------------|---------|------------|----------------------------------|----------------------|
| 1 | "Unemployment fell to 3.4%" (para 2) | Accurate with caveats | High | BLS Employment Situation, Jan 2023 release, table A-1 | Specify month; rate later revised |
| 2 | "(Smith et al., 2021)" (para 5) | Fabricated | High | No match in Crossref, Scholar, or journal index | Remove or replace with a real source |
Method notes: sources unavailable, paywalls, assumptions.For drafts, add a suggested correction for every non-accurate claim, worded to what the evidence supports.
Gotchas
- Multiple outlets repeating a claim is not corroboration if they share one origin; trace them to the source.
- Search snippets and AI summaries misquote; never cite a snippet.
- A fact-checker’s verdict is a secondary source; read their evidence and link the primary source.
- Satire and parody sites get screenshotted without attribution; check the origin.
- Statistics are revised; cite the release date and note revisions.
- Your own prior beliefs about a claim are the biggest risk; apply the same evidence bar to claims you expect to be true.
About this skill
Written and maintained by the Claude Skills editors to the same editorial standard as every skill in the directory: a selective trigger, a real workflow, known pitfalls and a verification step. First published , last updated . MIT licensed.