check-info-quality

name	check-info-quality
description	Flag stale or misleading content.
user-invocable	true
allowed-tools	["Bash","Read","Grep","Glob","WebFetch","Edit","Write"]

check-info-quality — flag stale, irrelevant, and misleading content

Three information-quality problems slip past both purge-hallucinations (fabricated references) and find-ai-tells (AI writing style), because each one requires judgment about a claim's relationship to something outside the claim itself — not just checking whether a reference resolves:

1. Out-of-date — the claim was true, but the world moved. Needs comparing the claim to its current state.
2. Irrelevant — the claim is true and current, but doesn't belong here. Needs judging the claim against its surrounding context.
3. Misleading / out-of-context — the claim is true in isolation, but misrepresents something (a source, a comparison, a scope) when read as written. Needs comparing the claim against what it's used to support.

All three are relational checks, not existence checks — the reason this is one skill with three passes rather than a purge-hallucinations variant.

When this fires

• "check info quality", "check-info-quality", "ciq", "audit this for information quality"
• "is this out of date / still accurate", "find stale information", "find outdated info", "does this cite the current version"
• "is this relevant", "does this belong here", "is this off-topic"
• "is this misleading", "is this out of context", "check this citation supports the claim", "does this quote/citation actually say that"
• Proactively, as part of a thorough content review (PR review, ardi cycle on documentation, or a grade-work pass) — fold in a pass of this catalog alongside find-ai-tells and purge-hallucinations.

Step 1 — Resolve the target and its context

Pick the narrowest target the user named: a file, a PR/MR diff (gh pr diff <n> / mcp__github__pull_request_read method: get_diff per tool-mappings.md in a remote session), or pasted text.

Unlike purge-hallucinations, this skill needs context beyond the target itself:

• For staleness — the current state of whatever the claim describes (a package's latest version, a tool's current API, a policy's latest revision).
• For relevance — the surrounding document (the chapter, the PR's stated scope, the section's topic) the claim sits inside.
• For misleading/out-of-context — the source a citation or quote points to, read in full, not just confirmed to exist.

Gathering that context is the expensive part of this skill; budget for it per finding, not just per target.

Step 2 — Run the three checks

A. Out-of-date information

For each factual claim that has a "current" value (a version, a status, a deprecation state, a guideline, a headcount, a "as of " statement):

1. Find the current value — check the package's release page / CHANGELOG, the tool's current docs, the policy's latest revision, or (matching check-dependency-updates' method) the relevant lockfile / manifest / uses: ref.
2. Compare. Flag when the claim's value differs from current and the claim doesn't already hedge it (e.g. "as of 2023," or "in v2," makes an old value correct, not stale).
3. Distinguish evergreen claims (a fact that doesn't change, a historical statement correctly framed as historical) from stale ones (an unhedged "current" claim that's since moved) — don't flag the former.

Grep starting points for likely staleness carriers: version-looking tokens (\bv?\d+\.\d+(\.\d+)?\b), date-anchored phrases ("currently", "as of", "the latest", "now supports"), and deprecation language ("deprecated", "no longer", "legacy").

B. Irrelevant information

For each passage, ask: does removing it change what the surrounding claim/task/section needs to establish? If the answer is no, it's a candidate.

• Off-topic tangent — a paragraph or aside that doesn't serve the section's stated purpose (a chapter on coding style digressing into deployment infrastructure).
• Scope creep — a PR/issue introducing changes unrelated to its stated goal.
• True-but-unrelated fact — a statistic or citation that's accurate but doesn't bear on the claim it's attached to (padding, not support).

Weigh relevance against the document's own stated scope — a section heading, a PR's title/description, an issue's acceptance criteria — not against a generic notion of what's interesting.

C. Misleading / out-of-context information

The check that most needs source-reading, not just source-existence:

1. Citation-claim mismatch — read the actual source the citation points to. Does it support the specific claim attached to it, a weaker version of the claim, or something else entirely? (Per shared/writing/citations.md — cite sources thoroughly means the citation must back the claim as stated, not just be topically related.)
2. Cherry-picked evidence — a true statistic or quote presented without the surrounding context that would change its reading (a result cited without its confidence interval, caveat, or contradicting follow-up finding in the same source).
3. False scope — a claim true for a narrow case presented as general ("X is faster" when the source measured one specific workload), or a comparison missing a relevant confound.
4. Missing context that reverses the takeaway — omitting a fact the reader would need to reach the same conclusion the author reaches.

This differs in kind from A and B: it requires reading the cited source's actual content, not just checking a link resolves (that's purge-hallucinations' job) or checking currency (check A) or placement (check B).

Step 3 — Report

One table per target, each finding tagged with which check (A/B/C) it falls under:

| Check | Finding | Location | Evidence | Severity | Proposed fix |
|-------|---------|----------|----------|----------|---------------|
| A: stale | "supports Node 14" | docs/setup.md:22 | package.json requires Node ^20 | blocking | update to current requirement |
| B: irrelevant | paragraph on CI caching | chapter3.qmd:80-95 | chapter scope is citation style | nit | move to a CI chapter or cut |
| C: misleading | "reduces errors by half" | README.md:5 | source reports 50% only for the largest of 3 cohorts tested | blocking | qualify the claim or cite the overall figure |

Severity mirrors the reviewer labels already in use in this corpus ("nit", "minor", "non-blocker", "optional") — see shared/workflow/address-every-comment.md for the treatment of each. Then, same discipline as find-ai-tells and purge-hallucinations: propose the fix, apply only on confirmation — never bulk-edit silently. A misleading-citation finding in particular needs a human call on how to reframe the claim; don't auto-rewrite it.

Custom agent for the detect phase

Steps 1-2 need no Edit/Write access. When the target is large (a whole PR, a full chapter, a batch of docs), delegate detection to a read-only agent so nothing gets modified before the report is reviewed — see agent-builder if this skill's fan-out step needs a dedicated persona, matching the pattern hallucination-detector and dependency-auditor already set for their skills.

Relationship to other skills

• purge-hallucinations — checks whether a reference exists; this skill (check C) checks whether an existing, real reference actually supports the claim it's attached to. Run both — a citation can pass purge-hallucinations (it resolves) and still fail this skill (it doesn't say what the text claims).
• find-ai-tells — checks how text is written; this skill checks what it claims. Complementary passes over the same target.
• check-dependency-updates (cdu) — the dedicated, deeper tool for check A when the staleness is specifically a pinned dependency/version; this skill's check A is the general-purpose, lighter-weight version for any "current" claim, not just dependency pins.
• check-rendered-refs (crr) — scans rendered output for broken crossrefs; this skill scans source content for quality, not render breakage.
• shared/writing/citations.md — the standing rule this skill's check C enforces: citations must actually back the claim.
• fact-check-prose (fcp) — the truth-side sibling: fact-check-prose verdicts a claim Accurate / Inaccurate (including re-deriving math and cross-checking rendered output); this skill's check C catches a claim that passes that verdict — it's true — yet still misleads through missing context, cherry-picking, or a citation that only weakly supports it. Checks A and B (staleness, relevance) have no fact-check-prose counterpart at all. Run both on a substantive prose review; neither subsumes the other. fact-check-prose also flags undefended claims (no citation, no internal reasoning, even if true) — a fourth failure mode distinct from this skill's check C, which only fires when a citation already exists but doesn't back its claim.
• address-every-comment / ardi — findings from this skill route through the same Address/Rebut/Defer discipline as any other review finding.
• shared/workflow/challenge-unnecessary-complexity.md — a sibling standing review rule, not a skill: flags unnecessary complexity or verbosity in prose, math, or code and asks for a simplification. Orthogonal to this skill's three checks — content can be maximally simple and still stale, off-topic, or misleading, and vice versa. Fold both into the same review pass.

Anti-patterns

• Flagging a correctly-hedged historical claim ("as of 2023...") as stale — hedged claims are evergreen by construction.
• Flagging check C from a citation's title or existence alone — read the actual cited content before judging mismatch.
• Judging relevance (check B) against a generic "is this interesting" standard instead of the document's own stated scope.
• Bulk-editing findings without proposing fixes first.
• Treating an unverifiable "current value" (paywalled source, private doc) as proof of staleness — report it as unverifiable, don't guess.