Fast Fact Check

Problem

Fact-checking often swings between two failure modes: an overlong multi-source report, or a fast answer that sounds confident but lacks citations. fast-fact-check targets verifiable factual questions. It pins down the exact claim or entity first, then searches and corroborates within a bounded budget. The goal is not exhaustive context; it is a bottom-line answer with confidence, evidence, and sources.

How it works

The skill runs under tiered budgets:
- simple: ≤2 min, 1 round, 1–3 parallel queries, ≤1 full-page fetch, ≥1 source.
- complex: ≤5 min, ≤2 rounds, 3–5 parallel queries, ≤3 full-page fetches, ≥2 independent sources.
It starts with a preflight check and rejects subjective or opinion questions that cannot be treated as facts. Then it decomposes the question into angles and issues the searches as parallel calls in a single message. It reads snippets first and fetches full pages only when a number, quote, or key judgment needs verification. Once the tier's source bar is met, it stops. Complex queries may run one additional targeted round; if evidence remains weak, the answer is marked uncertain.
The final output uses a BLUF structure: Answer, Confidence, Tier, 2–4 cited evidence bullets, numbered Sources, and optional Caveats. check_answer.mjs can validate citation resolution, confidence labels, and the per-tier source bar.

Boundaries

It fits quick verification of a fact, metric, event, version, or quote. It is not for comprehensive multi-source reports; use deep-research for that. Speed is bounded by traceability: below the source bar, the skill should not emit a guessed high-confidence answer, but report Low confidence or “could not confirm”.