Verify

Report what the piece claims and whether the text supports it. This skill flags; it never deletes or weakens a claim silently — an unsourced true statement is not a false one, and deletion is the user's call, not the skill's.

1. Extract every claim of fact: numbers, names, dates, quotes, superlatives (best, first, #1), and causal assertions (causes, drives, leads to). Skip opinion and clearly signposted judgment — a labeled opinion is not a claim to source.
2. Classify each: sourced (the text carries the source — quote it verbatim), unsourced (checkable, no source in the text), or unverifiable from text (a judgment dressed as fact, or checkable only externally). No external search unless the user authorized it; until then, "unverified" is the finding — never assume true, never assume false.
3. Rate by stakes, not difficulty: blocking (a named person or company, a public number, a superlative — publication risk), should-source (a claim that would embarrass if challenged), fine. Stakes govern severity, not how hard the claim is to check.
4. Report a table — claim · classification · source-or-none · severity — then the blocking set alone. For each unsourced claim, propose a source to find or a safe rewording; never a deletion the user didn't ask for.
5. Offer the external check of the unsourced set as a next step, naming what kind of source would settle each claim. Until the user opts in, the finding stands as unverified, and the draft should not go out carrying it as fact.