fact-check-prose — verify claims, reasoning, and computed values in prose
Prose review usually stops at style. This skill goes further: it checks whether the prose is actually true and its logic actually holds — per the standing rule in Morrison-Lab/ai-config’s shared/writing/fact-check-prose.md, which is the canonical statement of this policy (edit that file, not this skill’s procedure, if the policy itself changes).
When this fires
- “fact-check this”, “check the math”, “verify this reasoning”, “check this proof”, “is this accurate”, “does this claim check out”
- As part of any PR/MR review where the diff touches narrative prose — docs, READMEs, lecture notes, textbook chapters, commit/PR descriptions making a substantive claim. Run it alongside (not instead of) style review (
use-preferred-style,find-ai-tells) and terminology review (challenge-ambiguous-terminology, folded intoard/ardi).
Procedure
1. Check (forked)
Run the check-prose-claims skill on the target the user named. It runs forked as the read-only prose-claim-auditor agent: it extracts the factual claims, reasoning chains, computed values and citations, runs the pre-flight readiness gate for cited sources, checks each unit against local sources, a re-derivation or the rendered output, and returns a | Claim/step | Verdict | Basis | table plus citations worth adding. The extraction rules, the gate and the check for each kind of unit live in that skill’s file, not here.
For a document with many independent claims, the main session can instead fan the units out to prose-fact-checker workers in a Workflow, one unit each; the forked agent has no Agent tool, so it checks every unit itself.
2. Offer to apply corrections
On request, fix the flagged inaccuracies and reasoning gaps in place with Edit, preserving the author’s intent and voice, the same posture as find-ai-tells’s apply step. Don’t silently rewrite without being asked; report first.
Why the check phase is its own skill
With context: fork, a skill’s whole body runs in the forked agent, so a skill cannot fork only one step (ai-config#914). Step 2 needs Edit, which the read-only prose-claim-auditor agent cannot provide, so the checks live in check-prose-claims and this skill keeps the corrections in the main session, as purge-hallucinations does with detect-hallucinations.
Relationship to other skills
check-prose-claims— this skill’s forked, read-only check phase (step 1); run it alone for a report with no edits.challenge-ambiguous-terminology— catches phrasing whose meaning is unresolved; this skill catches claims and reasoning that are resolved but wrong. Run both during review.use-preferred-style/find-ai-tells— style and AI-tell scans; this skill is the accuracy/reasoning counterpart, not a replacement.ard/ardi— when a reviewer (human or bot) flags a factual or reasoning error,ardhandles the response disposition; this skill is how you (as reviewer) find that class of error in the first place.prose-fact-checkeragent — checks one unit handed to it, for fanning claim-checking out across aWorkflowwhen a document has many independent claims to verify at once.check-info-quality(ciq) — this skill verdicts a claim Accurate / Inaccurate / Undefended;ciq’s check C catches a claim that’s accurate by this skill’s own standard yet still misleads (missing context, cherry-picking, a citation that only weakly supports it).ciqalso covers two failure modes this skill doesn’t: stale claims (check A) and irrelevant content (check B). Run both.
Anti-patterns
- Don’t accept a claim because it sounds authoritative — check it.
- Don’t skip a derivation step because “it’s probably fine” — re-derive it or flag it as unverified, not silently pass it.
- Don’t trust the prose’s own description of a computed value when a rendered artifact (preview or local render) is available to check it against directly.
- Don’t limit citation suggestions to claims already flagged as uncited — suggest them anywhere they’d help a reader, per the fragment’s policy.