What problem does it solve? Documents accumulate outdated versions, wrong statistics, and misquoted sources, and manually verifying every claim against primary sources is slow and error-prone. This Skill decomposes a document into atomic claims, verifies each against retrieved evidence, and edits nothing without your explicit approval. ## Core Features & Use Cases - Claim-by-claim verification: Decomposes text into atomic claims, routes each by type (spec, version, date, statistic, computation, quotation) to the right source tier, and records a verdict with verbatim quotes, URLs, and access dates. - Evidence-first reporting: Renders a report where evidence precedes each verdict, with tiered approval (batch, per-item, or never-auto) so you judge corrections before any edit happens. - Safety invariants: Never verifies from memory, treats fetched pages as untrusted data with injection defense, and requires two independent sources for any proposed correction. - Use Case: Point it at a README or technical blog post before publishing; it flags an outdated dependency version, a wrong release date, and an unsupported performance statistic, then applies only the corrections you approve. ## Quick Start Fact-check the claims in docs/guide.md and propose corrections for anything outdated or contradicted.