Fact-Check Pass Skill

Everything in a published piece is a claim someone can be wrong about — a date, a number, a quote, an implication. This skill does what a good fact-checker does: it pulls every checkable assertion out of the prose, verifies each against a source, and stops the high-risk ones (stats, quotes, anything defamatory) from going out on a "probably right."

Working from a brief

Given the draft, produce the claim inventory and status — check what you can against the provided/known sources, and for anything you can't verify, mark it clearly rather than guessing. Never assert a fact is confirmed without a basis.

Required Inputs

Ask for (if not provided, else infer and label):

• The draft (article, script, report)
• Available sources — the reporter's notes, documents, links, transcripts
• Risk level / venue — where it publishes and how litigious/high-stakes the subject is

Output Format

Claim inventory

Every checkable assertion extracted:

| # | Claim (as written) | Type | Status | Source / basis | Fix | |---|---|---|---|---|---|

• Type: number/stat · quote · name/title · date · causal/implied · characterization.
• Status: ✅ confirmed · 🟡 needs sourcing · ⚠️ unverifiable · 🔴 wrong/misleading.

High-risk flags

The claims that must not publish unverified — statistics, direct quotes and their attribution, named individuals (especially any allegation), superlatives ("first," "only," "largest"), and anything potentially defamatory. Each with what's needed to clear it.

Fixes

For every 🔴/⚠️: the correction, a hedge to the supportable version, or "cut it." For 🟡: the source to obtain.

Sign-off

What's still open before this can publish, and a note that quotes should be confirmed with the speaker/transcript and legal review sought for anything defamatory.

Quality Checks

• ] Every checkable claim is inventoried, not just the obvious ones
• ] Each claim has a status and a named source/basis (or is marked unverifiable)
• ] Numbers, quotes, names, and superlatives are treated as high-risk and flagged
• ] Potentially defamatory statements are flagged for verification and legal review
• ] Nothing is marked confirmed without an actual basis
• ] Every problem claim has a concrete fix (correct / hedge / cut)

Anti-Patterns

• Reading for tone instead of extracting discrete claims
• Marking something "confirmed" on memory or vibes
• Letting a round number or a paraphrased quote through unchecked
• Missing implied/causal claims because they're not stated as facts
• Waving through a defamatory line without verification or legal input
• Rewriting the prose instead of flagging and sourcing the facts