Compatible with Claude, ChatGPT, Gemini, and Claude Code.
Fact-Checker
Checks every claim against primary sources.
Free, no account, no email — the command is all you need.
Already have it?
Nothing to install if it's already active
If this skill is already added in your Claude, there's no need to install it again. Just ask for what you want to do.
Not sure? Ask Claude: "which skills do you have available?"
“Use the fact-checking skill to verify the claims in this text against primary sources: …”
1 · The simplest way
Copy and paste into any AI chat
This button copies the whole skill. Paste it into Claude, ChatGPT, or Gemini, then describe what you need. No installation required.
In a chat, Word or Excel files can't always be generated: you'll receive the document as structured text instead, and Claude will flag it.
2 · Command line
/fact-checker installs in 3 steps
For Claude Code. The skill installs on your computer and stays available across all your projects.
Prerequisites: Claude Code + Node.js
The skill runs inside Claude Code, Anthropic's AI assistant installed on your computer. Node.js is the engine that runs the install command. Both are free.
Open PowerShell
Click Start, type "PowerShell", then press Enter.
Paste the command
Right-click inside the PowerShell window (or press Ctrl + V) to paste the command below, then press Enter. Wait a few seconds.
npx educasium --skill fact-checker --globalWhat you should see:
npx educasium --skill fact-checker --global
Fetching https://raw.githubusercontent.com/Chappygo-OS/claude-meta-skills/main/dist/fact-checker.zip
…
Installed "fact-checker". Restart Claude Code to pick it up.
Restart Claude Code
In Claude Code, type /restart (or quit and relaunch Claude Code). The skill now appears in the command list.
Don't use a terminal?
Using Claude.ai in your browser?
The assisted install works through a .zip file imported into Claude.ai. It takes longer and asks for your email.
What the skill does
View the source codeIn detail.
Pulls every verifiable claim from your text and checks each one against primary sources. Returns a verdict per claim: True, Mostly true, Mixed, Unverifiable, False or Outdated.
When to use it
You're about to publish AI-generated content, a LinkedIn post, a newsletter, or a sales deck.
Example usage
/fact-checker Check this article...In depth
How this skill really works
Fact-Checker pulls every verifiable claim from a text — statistic, named fact, quote, causal claim, technical assertion, or comparative claim — and checks it against external sources, then returns a structured report with a verdict per claim: ✅ True, ⚠️ Mostly true, 🔶 Mixed, ❓ Unverifiable, ❌ False, or 🔄 Outdated. Three use cases: checking AI-generated content before publishing, checking your own drafts, or fact-checking someone else's article or post.
Verification runs on a 4-tier source hierarchy — primary sources (official documentation, government data, peer-reviewed research) get priority, then high-quality secondary sources, then medium-quality sources that need cross-referencing, and finally unreliable sources never cited as proof (social media, AI-generated content, anonymous forums). Every high-priority claim must be cross-referenced against at least 2 independent sources.
The skill also specifically flags AI's typical fabrication patterns — the invented statistic, the phantom study, the fabricated quote, the outdated fact presented as current, the invented technical specification — documented in references/anti-patterns.md, and applies domain-specific verification guidance (AI/tech, business, science/health, law, history) documented in references/domain-guides.md.
When to use it
- Right before publishing AI-generated content — a LinkedIn post, newsletter, sales deck.
- To check your own drafts and outlines before developing them into final content.
- To check someone else's article, post, or deck before citing or sharing it.
- When a text contains statistics, attributed quotes, precise dates, or causal claims.
- When content was generated from a source document (transcript, notes) and needs checking for consistency against it.
- Embedded in an existing writing skill, to screen every draft before final output.
When not to use it
- Not for judging opinions, predictions, or explicitly acknowledged estimates — the skill ignores these categories; it only verifies what's presented as fact.
- Not for stripping AI-writing tells from a text — that's the role of /humanizer, to run after the factual check.
- Not for restructuring a prompt that happens to contain factual claims — /prompt-master handles the prompt's own structure.
- Not for open-ended investigation beyond checking the claims already present in the text — route to /deep-research for deep research.
The method
The 4-tier source hierarchy
references/source-hierarchy.md ranks sources into 4 authority tiers. This ranking — not a subjective sense of credibility — determines which sources can serve as proof and which need cross-referencing.
- Tier 1 — Primary sources
- Official company announcements, government and regulatory databases, peer-reviewed research, official documentation (APIs, RFCs, ISO/IETF/W3C standards), legal documents, direct unparaphrased statements. Use these first, always.
- Tier 2 — High-quality secondary sources
- Major news organizations with editorial and correction policies (NYT, WSJ, FT, Reuters, AP, Bloomberg), analysts with transparent methodology, Wikipedia for well-established facts only (a starting point, never an endpoint — check its cited sources).
- Tier 3 — Medium-quality sources, cross-reference
- Trade publications and industry blogs, company-sponsored research (possible selection bias), analyst reports without disclosed methodology, news aggregators. Always cross-reference against a higher tier.
- Tier 4 — Unreliable sources, never cited as proof
- Social media posts (even verified ones, unless it's the person's official statement), AI-generated content (circular verification is not allowed), anonymous blogs, forums, aggregator articles with no identifiable original source.
Steps
Extract every verifiable claim
Read the full text (fetch the page if a URL is supplied) and list every statistic, named fact, quote, causal claim, technical assertion, historical claim, or comparative claim. Ignore opinions, predictions, and acknowledged estimates.
Prioritize by risk
Rank each claim as High priority (numbers, quotes, causal claims), Medium (named facts, dates, technical assertions), or Low (widely known facts). Check High priority first; stop at Medium if time is limited.
Verify against primary sources
Search for a primary source first; cross-reference at least 2 independent sources for every high-priority claim. For fast-moving topics (AI tools, APIs, funding rounds), check the most recent source — information older than 12 months may be outdated.
Classify each claim
Assign a verdict: ✅ True, ⚠️ Mostly true, 🔶 Mixed, ❓ Unverifiable, ❌ False, or 🔄 Outdated — and document the source and publication date behind each verdict.
Document corrections
For every non-true claim: quote exactly what the text says, what the evidence actually shows, why the discrepancy exists (outdated, misattributed, simplified, fabricated), and a safer suggested rephrasing.
Render the structured report
Produce the report in the required format (summary, per-claim verdict, sources) followed by a confidence assessment (how many claims were verified with high confidence, lower confidence, or could not be verified).
Rules
- Ignore opinions, predictions, and acknowledged estimates — only verify what's presented as fact.
- Cross-reference at least 2 independent sources for every high-priority claim; a source that simply cites another one doesn't count as a second source.
- Treat every specific claim from AI-generated content as unverified until proven otherwise — including Claude's own output.
- Always render the report in the required format: verdict, exact quote, source and date for every claim.
- Mark ❓ Unverifiable rather than ✅ True for a quote whose primary source could not be found.
- Check source freshness for any fast-moving topic (AI, tech, funding) — a source older than 12 months may be outdated.
Mistakes to avoid
- Marking a quote attributed to a real person as ✅ True without having found its primary source.
- Counting two articles that both cite the same original study as two independent sources.
- Using Wikipedia as final proof instead of a starting point whose cited sources get checked.
- Using AI-generated content to verify other AI-generated content — circular verification is not allowed.
- Treating a rounded, uncited statistic ("50% of companies...") as plausible without searching for the original study.
- Letting an outdated claim (a figure accurate 3 years ago) pass as current without checking its freshness.
What this skill doesn't do
- Does not judge opinions, predictions, or acknowledged estimates — only claims presented as fact get verified.
- Does not strip AI-writing tells from the text — that step stays with /humanizer, to run after the factual check.
- Does not restructure the prompt containing the claims to check — /prompt-master handles the prompt's own structure.
- Does not run open-ended investigation beyond the claims already present in the text — routes to /deep-research for broader research.
- Does not treat AI-generated content (including its own) as a valid source to verify other content.
Skill contents — 4 fichiers · 25.5 KB(SKILL.md, references/anti-patterns.md, references/domain-guides.md, references/source-hierarchy.md)
Download for free