| name | verify |
| description | Aggressively fact-check claims in the last assistant message or specified text/file against authoritative sources (AWS docs MCP, WebSearch, WebFetch, agent-browser, gh, package registries). Depth-first per claim — exhaust evidence, follow references, run adversarial counter-queries before moving on. Triggers: "verify", "ファクトチェック", "エビデンスチェック", "本当に合ってる?", "ソースは?", "検証して", or right after producing technical content with API names, versions, config keys, quotas, ARNs, or URLs. |
Read-only, paranoid, depth-first. Exhaust evidence per claim before moving on. Accuracy > speed.
Input $ARGUMENTS: empty → last assistant turn (ask user to paste if unreadable); file path → Read; quoted text → use as-is.
Tiers — t1: RFC/W3C/ISO/AWS API ref/vendor official docs at current version/upstream README at pinned commit. t2: vendor blog/release notes/changelog/GitHub releases. t3: high-vote recent SO, established community docs. hint (never evidence, only a dig signal): AI Overview/AI summaries/low-vote Q&A/marketing.
Routing (primary + independent secondary; if primary unavailable, promote secondary, add a new independent one, note in report). AWS: mcp__aws-docs__aws___{search,read}_documentation → AWS blog/release notes via WebFetch. Static URL: WebFetch (+ web.archive.org when freshness matters). JS/SPA/anti-bot: Skill(agent-browser) + underlying JSON via WebFetch. Login-gated: Skill(agent-browser-9222). General: WebSearch → WebFetch top hit + independent second. GitHub: gh + WebFetch. Package version: registry CLI (npm view/pip index versions/cargo search/gem info) + upstream changelog. Default WebFetch; escalate to agent-browser only when insufficient; 9222 only when session required.
- Resolve target. Never verify text not read this turn.
- Extract every checkable claim exhaustively (API/IAM/ARN/quota/default/signature/flag/config key/env var/version/URL/spec citation/behavioral claim/deprecation/well-known paths). Exclude opinions, plans, subjective phrasing, accepted code. Order by severity: numeric constants > API/flag signatures > URLs/paths > behavioral. If >30, list all and ask user to prioritize.
- Across claims primaries may parallelize; within a claim primary → secondary → references → adversarial is strictly sequential.
- Dig recursively — when a source links to the document that defines the claim, open it and confirm the exact sentence. Chase until t1, no new references, cycle detected, or 5 hops per claim (then surface as
likely with depth note).
- Capture exact sentence/code block, anchor, verbatim value in the source's language (translation only as gloss). URL alone is not evidence. Search snippet is not evidence — open the page. Two sections of one source disagreeing = contested within that source; seek a tiebreaker.
- Freshness: confirm doc covers implied version. Stale doc vs newer release notes →
contested, side with newer. Re-fetch date-sensitive values in this session; never reuse prior-turn results.
- Adversarial (mandatory). Pick ≥3 from this menu (more if version-sensitive), one query each: (a)
"<X>" deprecated|removed|breaking change; (b) changelog diff implied-version vs current; (c) GitHub issues reporting opposite behavior; (d) alt docs (region/version/beta vs GA); (e) "X does not Y" / "X bug". All must return no counter-evidence to pass.
- Forced +1 drill-down on any of: version-dependent wording ("current"/"latest"/"as of");
deprecated/legacy/will be removed; numeric disagreement; summary-style quote without committed value; source covers only V−1 or V+1; second source is t3 while a second t1/2 exists.
- Confidence:
verified = ≥1 t1/2 pinpoint + ≥1 independent second (t≥3) same value + adversarial clean + all relevant refs followed. likely = t3-only, partial adversarial, or hop-budget hit. contested = two t1/2 give different values. mismatched = a t1/2 source contradicts the claim. t3-only caps at likely. Effort is not termination. For the assistant's own output, require two independent t1/2 sources on any plausible-looking API/flag/default/quota/version constant.
unverifiable only for hard blocks: paywall; login wall + no 9222; rate limit + retry exhausted; 404 + no archive; anti-bot on every viable path. Slow ≠ unverifiable. Never silently downgrade.
Report in user's language (default Japanese). Per claim, separate lines: - <claim> / 引用: "<exact sentence, source language>" / 根拠 (tier N): <URL> / 追認: <URL> / 反証: <H1>/<H2>/<H3> → clean. contested → replace 根拠/追認 with source A:/source B: (quote+URL+tier) + 推奨:. mismatched → add 正: <value> + 箇所: <file:line|直前の応答>. unverifiable → step-10 reason. Append URLs as <https://...>. Never edit the original claim.
Example — - SQS FIFO per-API throughput limit / 引用: "300 send, receive, or delete API actions per second" / 根拠 (t1): <url> / 追認 (t2): <url> / 反証: deprecated? / changelog diff / opposite reports → clean.
