科研 Skill

知识库内容一致性审查

semantic-wiki-review

Audits a research wiki’s content for contradictions, outdated claims, and weak support. Use for a meaning-level content review distinct from structural linting.

发布者 Codar安全审核通过v0.1.0更新于 2026-11-18

SKILL.md

Semantic Wiki Review

scripts/lint-wiki.py proves the wiki is structurally well-formed: frontmatter present, wikilinks resolve, no orphans. It cannot tell you whether two source pages contradict each other, whether a synthesis still holds after a newer source was ingested, or whether a claim lacks citation backing. That work needs reading — and that is what this skill is for.

Announce at start: "Using semantic-wiki-review to audit the wiki for content issues."

Not a gate. Manual trigger only. Runs alongside the workflow, never blocks it.

When to use

  • User asks: "are the wiki pages consistent?", "any contradictions?", "which synthesis pages are stale?"
  • After a batch of new sources was ingested and earlier syntheses may be stale.
  • Before promoting a status: review page to status: stable.
  • As a recurring sanity pass (monthly, before manuscript submission).

NOT for: structural lint (use wiki-lint), single-page proofreading, or fact-checking against external sources (that is a literature task).

Scope

Pick a scope before reading — full wiki audit is expensive. Defaults:

ScopeWhen
All status: stable pagesBefore manuscript submission
All pages touching topic XAfter ingesting a major new source on X
All syntheses citing source YAfter source Y was substantially revised
Recently-changed pages (last N days)Periodic check

State the scope to the user before reading.

Checklist

  1. Determine scope — confirm with user if ambiguous.
  2. List in-scope pages — read each fully (no skim).
  3. Build a claim ledger as you read — one line per substantive claim, tagged with [[source-key]] and page-of-origin.
  4. Flag issues in the categories below. Each has a recording channel — where the finding is written so downstream skills (lint, drafting) can act on it:
    • Contradictions — two pages assert incompatible things. → Record as a typed relations: edge with type: contradicts on the pages themselves (both directions), confidence: inferred unless a verbatim quote grounds it, plus a because. This is NOT a review_flag (a flag is about one page's own content; a contradiction is a relation between two).
    • Stale syntheses — a synthesis page relies on sources that a later source has complemented or refuted. Later means later published (the year in the bibkey, or orig_year), not later ingested: a 1996 paper ingested last week is still the older position. Start from the SUPERSEDED-CITED lines of lint-wiki.py's Research chronology section. → review_flags kind: stale.
    • Flattened chronology — a page sets positions from different years side by side as if all were held today: an older view in the present tense with no word about what came after, or a researcher cited for a view they later gave up (a revises edge exists, or a later source by the same author on the page says otherwise). → review_flags kind: stale, with a detail naming both sources and their years. Where the later source's answer is only in prose, also write the missing responds-to / supersedes / revises edge on the later page (confidence: inferred unless a quote grounds it). Later is not the same as right: flag the flattening, do not decide the debate.
    • Misdirected edges — ANACHRONISM lines in the lint output: a stance edge from an earlier source to a later one. → Move it to the later page with the fitting type; keep the because.
    • Unsupported claims — assertions without a [[source]] link or @citekey in the text. → review_flags kind: weak-support (or missing-citation when a specific citation is simply absent).
    • Overstatement — a claim stronger than the source supports (superlatives, precise figures without consensus). → review_flags kind: overstatement. This includes over-generalisation, the most frequent failure of model-written summaries: a source's hedge (probably, suggests) or scope (one site, stratum, period, corpus, sample) has fallen away, so the wiki states as general what the source stated as local or tentative. Compare the claim with its cited page; name the dropped qualifier in detail.
    • Generic conclusions — a synthesis conclusion that would be just as true of a neighbouring topic (swap the site, text or period and it still holds): a complex picture emerges, plays an important role, further research is needed, or a conclusion with no concrete anchor and nothing that could make it wrong. See executing-research-plan → Writing a synthesis. → review_flags kind: weak-support, detail starting with generic: and quoting the sentence. (Weak support, because a claim that names no evidence cannot be checked against any.)
    • Missing cross-references — pages discussing the same entity / concept without linking each other. → Report only (a linking suggestion, not a content defect); the user or a follow-up adds the wikilink.
    • Suspect / aggregator citations — citations whose source URL points to aggregator hosts (academia.edu, researchgate.net) or sources flagged as suspect/pseudoscientific. → Report only. See MCP Optimisation below.
  5. Record structured findings on the affected pages — for the content categories above, append a review_flags entry to the flagged page's frontmatter (see "Recording review flags" below), and add/strengthen the relations: contradicts edges for contradictions. This is the ONLY editing the audit does: frontmatter metadata only, never the page prose. Existing open flags on a page are left in place (dedupe by kind + detail); do not silently reopen a state: resolved flag.
  6. Write a review report to knowledge/_meta/semantic-review-<YYYY-MM-DD>.md (frontmatter: type: synthesis, status: draft, author: llm) — the human-readable digest of the same findings, plus the report-only categories.
  7. Never touch page prose or status. The flags are advisory metadata; the user (or a follow-up skill invocation) decides which issues to fix, and only the user promotes status. Resolving a flag means setting its state: resolved (with a resolved: date) once the concern is addressed — not deleting it.

Report format

---
title: "Semantic Wiki Review — <scope> — <date>"
type: synthesis
created: YYYY-MM-DD
updated: YYYY-MM-DD
status: draft
author: llm
---

## Scope
<one sentence>

## Pages audited
- [[page-1]]
- [[page-2]]
- ...

## Contradictions
| Pages | Issue | Suggestion |
|-------|-------|------------|
| [[a]] vs [[b]] | A claims X, B claims ¬X | Resolve in synthesis or note dispute |

## Stale syntheses
| Page | Outdated by | Why |

## Flattened chronology
| Page | Earlier position (year) | Later position (year) | What the page does |

## Unsupported claims
| Page | Claim | Suggested source / mark as opinion |

## Missing cross-references
| Page | Should link | Why |

## Suspect / aggregator citations
| Page | URL | source_class | Primary source known? | Suggested action |

## Summary
<2-3 sentences: how healthy is the wiki, what to prioritise>

Recording review flags

Single-page content findings are written to the flagged page's own frontmatter as a review_flags entry (schema: schema/knowledge-frontmatter.schema.json). This is a distinct axis from status (human-owned maturity — never touched here) and from relations: contradicts (page↔page conflicts). A page can be status: stable and carry an open flag at the same time — that is exactly the case worth surfacing: a blessed page a newer source now undercuts.

# appended to knowledge/synthesis/chronologie-debatte.md frontmatter
review_flags:
  - kind: overstatement          # overstatement | weak-support | stale | missing-citation | open-question
    detail: "Dates the destruction 'securely' to 925 BCE; the cited source says 'probably'."
    raised_by: semantic-wiki-review
    detected: 2026-07-03
    state: open                  # open | resolved

Why on the page and not only in the dated report:

  • It gates drafting. drafting-manuscript refuses to draft from a page with an open flag without a logged override — the finding cannot be silently skipped. A detached report can.
  • It travels with the page — visible in Obsidian and to any later reader, not buried in a _meta/ file dated weeks ago.
  • It is auditable and resolvable in place — state: resolved + a resolved: date closes it while keeping the trail; wiki-lint reports the open ones.

The dated report (step 6) stays as the human-readable digest and is the only place the report-only categories (missing cross-references, suspect citations) live. Findings and report are two views of the same audit — keep them consistent.

Web-source classification

Classify every cited URL per docs/source-classes.md. (Until v0.45 the optional dao-searxng-mcp did this; it is discontinued — the labels are unchanged.)

Procedure:

  1. Collect every external URL from the audited source / synthesis pages (frontmatter sources: or body links).
  2. Per URL: fetch the page (WebFetch / webfetch) and assign its source_class (primary_publisher | academic_repository | preprint_server | aggregator | suspect | grey_lit_or_unknown) by the rubric in docs/source-classes.md.
  3. URLs with source_class=aggregator or suspect go into the new "Suspect / aggregator citations" issue category, with a note on whether the primary source (DOI) is known.
  4. No automatic fix — the report is advisory; the user decides replacement or retention with rationale.

Red Flags

ThoughtReality
"I can check this in my head"Above 5 pages impossible — keep a ledger.
"The linter already checked"The linter checks structure, not content. Different job.
"I'll fix issues as I find them"No — the audit repairs nothing. It records review_flags (frontmatter) and writes the report; fixing prose is a separate step.
"Stable means correct"Stable means human-reviewed at time X. It can still be stale — flag it, don't demote it.
"A contradiction is a page flag"No — a conflict between two pages is a relations: contradicts edge; a review_flag is about one page's own content.

Key Principles

  • Metadata-only, never prose. The audit records review_flags and relations on pages and writes the dated report — it never edits page prose and never changes status. Prose repair is a separate, later step.
  • One report per run, dated, in knowledge/_meta/; flags live on the pages.
  • Scope before depth. 10 pages honestly read beats 100 skimmed.
  • No gate. This is a tool, not a precondition for other skills.

示例提示词

帮我检查知识库里有没有互相矛盾的论断或被新文献推翻的旧结论。

所属插件

人文学科面向历史、哲学、语言学、考古与艺术研究:TEI/OCR 与数字档案、论证分析、语料库与语音学、放射性碳测年、乐谱与音频分析,以及人文论证写作与审稿。

安装该插件会一并获得同领域的 62 个 Skill / MCP。

同插件的其他科研 Skill

常见问题

知识库内容一致性审查能做什么?

Audits a research wiki’s content for contradictions, outdated claims, and weak support. Use for a meaning-level content review distinct from structural linting。

如何安装知识库内容一致性审查?

在科搭 Codar 左侧点击「插件」,在「组件」页签搜索「知识库内容一致性审查」并安装;在 Claude Code 或 Codex 中可以发送本页的安装提示词,或下载 Skill 包后解压到 ~/.claude/skills 或 ~/.codex/skills。

知识库内容一致性审查需要哪些权限?

该科研 Skill没有声明额外权限;执行具体操作时仍由 Agent 按自身规则请求确认。

知识库内容一致性审查经过官方审核吗?

是。SpeedSkills 收录的每个版本都会记录内容摘要与权限声明,当前版本已通过安全审核。