The factcheck Skill — Structured Claim Verification Before Publication

Publishing a factual error is expensive to correct and impossible to fully retract. The factcheck skill runs a structured verification pass over an article or social post before it goes out, catches incorrect and outdated claims, and applies the minimum correction needed — without touching the author's phrasing.

The skill is invoked with a file path as its argument. It works through six sequential phases and keeps intermediate artefacts in a timestamped directory under /tmp/factcheck-{timestamp}/, so every step is auditable and the original file is never modified without a backup.

Claim extraction and importance tiers

Phase 1 extracts every verifiable factual claim and assigns each one a type and an importance level. Types cover statistics, dates, names of people and organisations, technical specifications, scientific findings, legal claims, and attributed quotes. Importance levels are critical, high, medium, and low — the level determines how many independent sources are required before a verdict is assigned.

Critical claims drive the article's central thesis and require at least three sources. High-importance supporting facts need two. Medium background context needs one. Low-importance common-knowledge claims get a spot check only when they look suspicious.

The verification chain

Phase 2 processes claims from most critical to least, using a Chain of Verification approach. For each claim, the skill formulates a specific search query, searches authoritative sources — official product pages, API documentation, official blog posts, peer-reviewed papers, established news outlets — and fetches the source page to confirm the claim against the actual text, not a search snippet. Cross-referencing a second independent source is required for critical and high-importance claims.

Each claim gets one of six verdicts: ACCURATE, INACCURATE, OUTDATED, MISLEADING, UNVERIFIABLE, or NEEDS_CONTEXT. A confidence score from 0.0 to 1.0 accompanies the verdict. Social media posts are explicitly excluded as authoritative sources. If a source cannot be found, the skill reports UNVERIFIABLE rather than fabricating a citation.

Corrections and consistency

Phase 3 checks internal consistency across the document — sections that contradict each other, timeline anomalies, ripple effects where one corrected claim changes the meaning of others. Phase 4 applies corrections: only claims that are not ACCURATE get touched, the author's sentence structure and tone are preserved, and each change is annotated with an inline HTML comment citing the source.

Phase 5 shows the author a diff summary and requests review before anything is written back to the original file. Phase 6 applies the approved changes, confirms the backup exists, and reports the total claims checked, corrections applied, and any remaining unverifiable claims that need manual attention.

What it does not do

The skill makes no stylistic edits beyond what a factual correction requires. It keeps the article in its original language — corrections and inline comments are in the same language as the text. It does not rewrite sections, restructure arguments, or change the author's voice.

Read more about the Datarim pipeline in what Datarim is, or see how the expectations-checklist skill applies a similar verification logic to task acceptance criteria.