跳至正文
OpenAgentSkill
浏览已开放 · 目录记录日期 2026-11-21 · 自动更新和账户服务暂时关闭。
其他技能

wiki-audit

Use when fact-checking a single wiki page against its cited sources — verifies that every footnote actually supports its claim and surfaces uncited factual claims. Run after ingesting a high-stakes page or any time you want confidence in one page's accuracy.

kfchou · 已有静态检查记录

安装与使用

运行命令前先阅读来源与依赖说明。复制命令不会自动安装或执行。

npx skills add kfchou/wiki-skills --skill wiki-audit

技能说明


name: wiki-audit description: Use when fact-checking a single wiki page against its cited sources — verifies that every footnote actually supports its claim and surfaces uncited factual claims. Run after ingesting a high-stakes page or any time you want confidence in one page's accuracy.

Wiki Audit

Verify a single wiki page against its cited sources. Two phases: detect uncited factual claims, then verify cited claims by dispatching one subagent per source in parallel. In strong mode (wiki-audit strong) a third phase adds a cross-model adversarial review: a different-provider model re-examines the same claims for overreach and contradiction, and disagreements with the normal pass become findings.

Pre-condition

Find SCHEMA.md (search from cwd upward, or ~/wikis/). If not found, tell the user to run wiki-init first. Read SCHEMA.md for the wiki root path and the Citations section (the rules and footnote format the audit enforces).

Link style. Read link_style: from SCHEMA.md. If the field is missing or <wiki-root>/config/link-style.md does not exist, default to obsidian ([[slug]]). Otherwise, read <wiki-root>/config/link-style.md. Footnote slug-targets in audited pages may be in either obsidian or markdown form — use the ## Parse rule to recognize them both. The audit report you write is a wiki page, so its own cross-references and slug-target examples emit using the ## Emit rule.

If SCHEMA.md has no Citations section (older wiki, initialized before this skill existed): use the fallback convention below for this run, and offer at the end to append the Citations section to SCHEMA.md so future operations stay consistent.

Cite every non-common-knowledge factual claim. Granularity is paragraph or claim,
never per-sentence. Format: Markdown footnotes. Two citation kinds:

Quote:     [^N]: <target> <locator> — "<verbatim quote>"
Synthesis: [^N]: <target> <locator> [synthesis] — <what supports the claim>

Three rules:
1. Target is one of: a slug reference to a source-type wiki page (form per the
   wiki's link_style; fallback `[[source-slug]]`), raw/<file> or assets/<file>
   (a local file path), or a URL. Never an entity / concept / analysis page.
2. A semantic locator is present (§section, p.N, [HH:MM:SS], URL anchor, dated
   post). PLUS a line-range token (L142-145, L142, or L142-145,L201-203) when the
   resolved raw source is text-addressable (markdown / plaintext / code / cached
   HTML). The line-range points into the raw source file and is required for both
   citation kinds. Exempt (semantic locator only): PDFs, transcripts, live URLs
   with no local copy.
3. Either a verbatim quote, or the [synthesis] tag plus a description of what
   the cited range supports.

Mode: if the invocation arguments contain the word strong (e.g. wiki-audit strong transformer-architecture), enable strong mode — run Phase C (§3b) after Phase B. Otherwise run normal mode (Phases A and B only). The remaining argument token, if any, names the page.

If the user did not name a page, ask which page to audit. Accept slug, filename, or absolute path. Resolve to wiki/pages/<slug>.md. Audit one page per run.

Process

1. Read the target page

Read the full page. Note:

  • The frontmatter sources: list (which sources the page formally cites).
  • All footnote definitions ([^N]: ...) and references ([^N] in body text).

If the page has zero footnotes but contains factual content, that is itself the audit result — every claim becomes a Phase A finding. Still run Phase A; skip Phase B.

2. Phase A — uncited claim detection (1 subagent)

Dispatch one subagent. Give it:

  • The full page contents.
  • The Citations section copied from SCHEMA.md.
  • The page's sources: list.

Task: list every non-common-knowledge factual claim that lacks a footnote. Return a structured list of (line number, claim text, suggested-source-from-the-sources-list-or-"unknown").

The subagent applies the SCHEMA.md "what to cite" rule: paragraph- or claim-level granularity, common knowledge exempt.

3. Phase B — cited claim verification (N subagents, parallel)

For every footnote definition in the page, parse:

  • The target — one of a slug reference (matched via ## Parse from config/link-style.md — extract the slug from either obsidian or markdown form), a path under raw//assets/, or a URL.
  • The semantic locator (§section, p.N, timestamp, URL anchor, dated post).
  • The line-range L<start>-<end> if present (text-addressable sources carry one).
  • Either the verbatim quote or the [synthesis] description.

Resolve each target to readable content:

  • slug reference → read wiki/pages/<slug>.md to find the raw file path (in its **Source:** line). Then read that raw file from raw/ or assets/.
  • raw/<file> or assets/<file> → read the file directly.
  • <URL> → check whether a cached copy exists in assets/ (filename derived from URL). If yes, read it. If not, mark the footnote 🚫 source-missing (do not re-fetch — that belongs in the fix step).

Any target that cannot be resolved gets verdict 🚫 source-missing; do not dispatch a subagent for it.

Fast path — footnotes carrying a line-range (L…): Do not dispatch a subagent. Resolve the target to its raw file and read only the cited lines (e.g. sed -n '142,143p' raw/<file> or the Read tool with that offset/limit). Then:

  • Quote footnote: string-match the verbatim quote against the text in that line range, ignoring leading/trailing whitespace and collapsing internal runs of whitespace to single spaces. Match → ✅ supported. No match → ❌ unsupported (note what the lines actually say).
  • Synthesis footnote ([synthesis] + L…): dispatch a subagent, but give it ONLY the cited line slice (not the whole source) plus the synthesis description, and apply the verdict rubric below. This is the bounded-read case — far cheaper than reading t
阅读完整来源 ↗