菜单
Lint the knowledge base at `~/.knowledge/` (LLM Wiki). Deterministic checks are delegated to a script; contradictions, broken links, and staleness are reported as proposals only (no auto-fix). Use for KB health checks, contradiction detection, right after a batch ingest, or before relying on the KB as evidence.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
エージェントの文脈に影響する永続層 (メモリ・CLAUDE.md・規約文書・指針文書・スキル文書等の日本語文章) の日本語品質を、`textlint-ja` と独自規則で再現可能に点検する。英単語混入等の文章規律違反を検出。ユーザーが「日本語品質チェック」「textlint かけて」「英語混入チェック」「メモリ点検」「`memory sanitize`」等を指示したら起動する。
Write or update your handover document so the next holder of this job seat can take over your responsibilities. The seat you hold is identified by (project, role) in the document's frontmatter; the document is a handoff report, not a personal diary. Use this skill whenever the user mentions 'handover', 'hand over', 'pass it on', or '/handover'. NEVER invoke this skill proactively, autonomously, or as part of completing a task — finishing work, wrapping up, ending a session, or saving progress are NOT triggers for this skill unless the user explicitly asks for a handover.
Take over a job seat left by a previous holder — adopt responsibility for the work going forward, audit the predecessor's claims before relying on them, and continue under your own name. Every referenced artifact becomes a mandatory read task, drained before any work starts. The predecessor's report is yours to audit, not yours to inherit unexamined. Use this skill whenever the user mentions 'takeover', 'take over', 'pick up where we left off', or '/takeover'. NEVER invoke this skill proactively or autonomously — resuming or continuing work is NOT a trigger for this skill unless the user explicitly asks for a takeover.
Retrospective grounded in the garbage-in-garbage-out principle: traces problems from Output back to their upstream origin across six stages, then fixes at the stage where the cause lives — knowledge deficiencies get knowledge fixes, not downstream rule patches. Implements fixes in the current session. Use after finishing a task, before creating a PR, or whenever the user says 'retrospective', 'wrap up learnings', or similar.
Edited a project's manifest, lockfile, or config? Added or deleted a whole project? Then its `.knowledge/` wiki page is now stale — use this skill to ingest (regenerate the page) or deingest (remove it). Assume the user won't mention the wiki or KB; trigger from the file touched or the action alone. Fire when: - a project's stack/dependency/build/deploy/structure-defining file changes — package.json, any lockfile (pnpm-lock, uv.lock, gleam.lock), pyproject.toml, gleam.toml, tsup.config, angular.json, wrangler.toml, Dockerfile/Cloud Run, pnpm-workspace.yaml, new workspace packages — or its CLAUDE.md; - a project is cloned, forked, or placed under a KB anchor (`~/works/`, `~/.claude/`), or a project directory is removed from one; - **AI-generated intermediate-artifact files appear, change, or are discovered** anywhere under a KB-covered repo — design docs (`docs/designs/*.md`), review records (`docs/reviews/*.md`), plans (`docs/plans/*.md`), acceptance-test records, retrospectives, conversation logs — **regard
Use this skill whenever the user is trying to recall, verify, enumerate, or reconcile facts about their own collection of projects — especially "which of my projects do X" questions. X can be a library, package, framework, version, service, cloud platform, deploy target, architectural pattern, or "what even is this project." Trigger on memory-checking and inventory intent: enumerating consumers/users of something, confirming a remembered list ("…だっけ", "…で正しい？", "…以外もあったよな", "…の気がする"), mapping relationships or dependency chains between projects, or sorting projects by which tech/version they use. Trigger even when phrased casually, even when one specific project or package is named, and even though the user never mentions a KB, wiki, or `.knowledge/`. The skill answers from the knowledge base at `~/.knowledge/` with citations. Do NOT trigger for: writing/debugging code inside a single named project, learning a technology in general, or adding/removing/maintaining KB pages.
| name | kb-lint |
| description | Lint the knowledge base at `~/.knowledge/` (LLM Wiki). Deterministic checks are delegated to a script; contradictions, broken links, and staleness are reported as proposals only (no auto-fix). Use for KB health checks, contradiction detection, right after a batch ingest, or before relying on the KB as evidence. |
Mechanical checks are delegated to a script (do not count things in-context).
kb-lint skill's kb-lint.sh. It enumerates:
type; reserved files (index.md / log.md) violating §6 / §7 structure. Any non-zero count is a hard violation of the spec.source_commit first, mtime fallback).wiki/projects/, and dead citations. Detection only — resolution happens at ingest time per the source-authority order.~/.knowledge/wiki/log.md. Do not auto-fix; report proposals only (human-in-the-loop gate).