shimo4228
UserCurated, MIT-licensed Agent Skills, subagents & rules from shimo4228's harness — open Agent Skills standard, not Claude-Code-only. Lift what you want.
Categories
Indexed Skills (111)
verify-bootstrap
repo に機械ゲート（format / lint / type check / security / dependency audit / test）を立てる、または既存のゲートが古びていないか棚卸しする。Use when starting a new project, when a repo has no automated quality gate, when the user says 「lint を入れて」「この repo にゲートを立てて」「型チェックを入れたい」「静的解析を整備して」「ツールが古い」「verify を棚卸しして」, "set up linting", "add a quality gate", "bootstrap the toolchain", or invokes /verify-bootstrap. 使うツールは skill が知っているのではなく、その時点で search-first に調べさせる — repo のスタックが何であれ同じ手順で回る。NOT for — 既に立っているゲートを 1 回実行するだけ（それは repo の verify entrypoint を直接実行）、ハーネス自身の設定監査（→ config-gc）、コードの意味的レビュー（→ implementation-chain の Review 群）。
task-triage
Run one cycle of the task-triage loop over a repo's task ledger — judge every open task (verify its premise in code, check its start condition against the 照合先, decide whether it is still worth doing, look for a better solution), then dispatch the accepted ones to fresh implementation sessions and act as their independent judge until the human merges. Use when the user says 「残タスクを見て」「タスクを整理して」「台帳を回して」「dispatch して」「未マージある？」, invokes /task-triage, or when a task ledger has grown and nobody can say what is dispatchable. This skill is the judgment layer of the loop (Fable = judge, Opus sessions = build, human = last switch); the vocabulary of ledger states is owned by task-stocktake and is not redefined here. NOT for consolidating scattered task files into a ledger (task-stocktake), NOT for deciding a single build-or-not question (architect), and NOT for running a task yourself — a triage session reads, judges, dispatches and verifies; it does not implement.
skill-creator
skill / agent 定義を新しく書く・大幅に改修するときの入口と草稿ゲート。著者が「skill 作って」「この手順を skill にして」「agent 定義を書いて」「この skill を書き直して」と言ったとき、learn-eval が Promote を決めたとき、hooks/skill-create-notice.sh が新規作成を検知したとき、/skill-creator で使う。intent を 1 packet に固定し、隣接 skill との境界を library 全体で引き、Fable 向けの書き方で書き、fresh-context の subagent に集計しない named verdict（Publishable / Fix / Drop）を出させ、著者通読で閉じる。NOT for — 既存 skill 群の棚卸し（→ skill-stocktake）、遵守率の測定（→ skill-comply）、参照切れ・所有権の検査（→ skill-health）、公開 repo への同期（→ harness-sync）、会話からの抽出と Save/Drop 判断（→ learn-eval）。
session-judgment-mining
過去の Claude Code セッション群（~/.claude/projects/<project>/*.jsonl）を遡及的に一括発掘し、ユーザーが繰り返し下した判断・価値観を抽出して skill / rule に正本化するワークフロー。人間発話の抽出 jq パターン、既存資産（skills / rules / ADR / memory）とのカバレッジ照合による重複回避、価値観リファレンス（why）と判断ゲート（when/what）の二層設計判断、既存スキルとの矛盾解消（免除条項）と memory への昇格マークまでを扱う。Use when — 「過去セッションから私の判断・価値観をスキルにして」「セッション履歴を紐解いて規約化して」、同じ指摘・修正がセッションを跨いで繰り返されていると気づいたとき、memory の feedback が溜まって確率的リコール頼みになっているとき。NOT for — 過去ログから記事の問いを発見 → session-theme-mining、現行セッションからの単発パターン抽出 → learn-eval、skill 品質の監査 → skill-stocktake、既存 skill 群からの rule 蒸留 → rules-distill、会話ログの要約・議事録作成。
wiki-harvest
研究 repo セッションから LLM wiki (Obsidian Vault wiki/concept/) を read-only で走査し、その repo の主担当 concept ページから「repo の次アクションを変えうる候補」だけを抽出して、一次出典付き・landing slot マップ付きのランク付き候補台帳 (ledger) を repo の .notes/ に生成する。Use when the user invokes /wiki-harvest, asks「wiki から repo に還元して」「wiki の有益分を AKC/AAP/CA/authorship に持ってきて」, or when closing the daily-research→wiki→repo loop. wiki への書き込みは行わない（それは vault セッションの /ingest）。chat 上の自由質問は wiki-query。
harness-boundary
agent 環境に mechanism（rule / skill / hook / agent / workflow / runtime 拡張 / prompt chain）を追加・変更・レビューするとき、それが 6 層（model capability / skill = 手続き記憶 / values・policy / eval / data・memory / runtime）のどこに属するか、なぜモデル自身に任せられないか、次のモデル世代で不要になるか、runtime を Claude Code → Pi → Codex と交換しても残す価値があるかを問い、Keep / Move / Simplify / Make temporary / Delete / Defer を返す設計レンズ。Use when — 「これはハーネスに入れるべきか」「どの層に置くか」「モデルに任せられないか」「runtime 変えても残るか」「harness が肥大している」「この hook / rule / workflow を足していい？」, when implementation-chain の Plan で harness 自体（~/.claude の rules / skills / hooks / agents / settings）を変更する task と判定されたとき, or /harness-boundary. Delete / Simplify は成功として扱う。NOT for — 未構築物の build-or-not 単体（→ agent architect）、設置済み資産の定期監査と Retire / Dissolve の verdict（→ rules-stocktake / skill-stocktake / agent-stocktake。本 skill は証拠を渡すだけ）、世代交代時の一括照合（→ generation-audit）、loop 構��の妥当性（→ loop-design-check）、harness の作り方の処方（→ agent-harness-construction）。
llms-txt-writer
AI 向けドキュメント（llms.txt / llms-full.txt、FAQ ページ、用語集等）を書くスキル。Answer.AI llms.txt 標準準拠と GEO/AEO 静的解析の両輪で、ChatGPT / Perplexity / Gemini に引用されやすい AI-facing コンテンツを生成する。audience が AI 専用の doc に使う（README 等の人間向けには使わない）。
agent-stocktake
Audit ~/.claude/agents/*.md (subagent definitions) for description-layer residency cost, body-layer quality, suppression instructions, staleness, and substrate absorption, assigning Keep/Improve/Update/Merge/Demote-to-skill/Dissolve/Retire verdicts. Use when the user says "audit my agents", "agent stocktake", "which agents should I retire or merge", 「agent を棚卸しして」「エージェント定義を見直して」, or when the model generation changed and agent bodies written for the previous one may suppress or over-constrain the current one. NOT for — skill quality → skill-stocktake; rules → rules-stocktake; runtime 層との横断照合 → generation-audit; whole-config GC → config-gc.
context-sync
Audit and fix project documentation — detect role overlaps between context files (CLAUDE.md, ADR, README, graph.jsonld), migrate misplaced content, check freshness against code, and create missing docs. One command to keep all project context healthy.
jsonld-knowledge-graph
Design and ship a companion JSON-LD knowledge graph (graph.jsonld) next to llms.txt for projects with stable concept-level structure. Encodes domain entities and relationships as schema.org triples for LLM citation. Use when project has matrix / hierarchy / phase-binding structure that prose alone leaves implicit, AND that structure is stable across releases. Defers llms.txt navigator wording to llms-txt-writer.
llms-txt-writer
AI 向けドキュメント（llms.txt / llms-full.txt、FAQ ページ、用語集等）を書くスキル。Answer.AI llms.txt 標準準拠と GEO/AEO 静的解析の両輪で、ChatGPT / Perplexity / Gemini に引用されやすい AI-facing コンテンツを生成する。audience が AI 専用の doc に使う（README 等の人間向けには使わない）。
readme-writer
README やプロジェクトのトップページ（repo を開いた人が最初に見る入口）を書く・直すときに使う。こんな時に呼ぶ — README が長い／継ぎ足しで文脈が重くなり初見で分からない、開いて数十秒で「何のプロジェクトで自分向けか」が伝わる入口にしたい、長い rationale・ADR 参照・内部史を docs/ に逃がしたい、研究・DOI repo の README を引用付きで読める長さにまとめたい、GitHub の About（description / topics）が README と食い違っている。証拠スクリプト（readme_evidence.py）+ fresh context の判定器（readme-judge）+ 上限 2 ラウンドの改稿ループ + review panel + 著者通読 GO で回す。短く・走査しやすくしつつ、LLM が README 一枚で要点を復元できる情報フロアは残す。CLI でも研究 repo でも、日本語でも英語でも、新規作成でも既存改善でも対象。AI 専用ドキュメント（llms.txt 等）は → llms-txt-writer、記事・エッセイは → writing-ecosystem。
release-doi
DOI-registered research repo (Zenodo) のリリース手順。README 多言語 / CHANGELOG / CITATION.cff / pyproject.toml / llms.txt / glossary を整合させてから tag push、Zenodo 自動採番後に新 DOI を反映し、Software Heritage archive + SWHID 記録 (intrinsic identifier 層) まで行う 4 phase + post-release ワークフロー。AKC / AAP / contemplative-agent など shimo4228 系の研究 repo で再利用する。
repo-asset-stocktake
Audit a project repo's non-code assets — tool configs, CI/GitHub workflows, runbooks, other docs — for assets whose consumer has vanished, and assign Keep/Update/Retire/Merge verdicts. Use when the user says "audit my repo assets", "which configs/workflows/runbooks are dead", "repo asset stocktake", 「非コード資産を棚卸しして」「使われてない設定/workflow/runbook を洗い出して」. NOT for — dead code → refactor-clean; doc-role overlap across CLAUDE.md/graph.jsonld/ADR/README → context-sync; ~/.claude config GC → config-gc; skills or rules → skill-stocktake / rules-stocktake.
jsonld-knowledge-graph
Design and ship a companion JSON-LD knowledge graph (graph.jsonld) next to llms.txt for projects with stable concept-level structure. Encodes domain entities and relationships as schema.org triples for LLM citation. Use when project has matrix / hierarchy / phase-binding structure that prose alone leaves implicit, AND that structure is stable across releases. Defers llms.txt navigator wording to llms-txt-writer.
skill-stocktake
Audit installed Claude skills for quality and surface Keep/Improve/Update/Retire/Merge verdicts. Use when the user says "audit my skills", "stocktake", "review my skills", "which skills should I retire or merge", "do a quality pass over my skills", or "/skill-stocktake". NOT for creating or improving a single skill (that is skill-creator) and NOT for whole-config GC across hooks/permissions/MCP (that is config-gc).
authorship-strategy
maker / 実践者が AI 時代の著者戦略を実地で探るための判断フレームワーク。あなた自身の DOI-registered idea-rescue repo 群（著者自身の例は AKC, Contemplative Agent）で適用する。core principle は「AI 時代のオーセンティシティ inversion」— 3 軸（scarcity → diffusion / exclusivity → derivation / enclosure → openness）の反転。主 audience は層別（ADR-0022 二層会計）— doctrine 層 = LLM-mediated channels（LLM 直接 + LLM 経由で情報を得る人間）、essay 層 = 同時代の human reader（同時代 KPI の観測+操舵可）。creative reuse > training > investigation の preference 階層。4 層 framework (authenticity → diffusion → idea/scaffold 判別 → tactics) で判断軸を提供。マネタイズ禁止、tool-agnostic、any-usage 容認、permissive license 原則
generation-audit
Model-generation-change audit orchestrator — collect the runtime layer (system prompt + tool descriptions) from the live session, cross-check every self-authored asset (rules / CLAUDE.md / skills / agents) against it, classify mismatches as conflict / redundancy / drift, judge each with the intent-evidence-freshness-expiry frame, then hand the evidence to rules-stocktake / skill-stocktake / agent-stocktake for verdicts. Use when a new Claude model generation ships and assets written for the previous one may now conflict with the substrate — 「世代交代したのでハーネスを照合して」「新モデルに合わせて棚卸しして」 "generation audit", "audit my harness against the new model". NOT for — routine single-layer audits (call the stocktakes directly); rule compliance → skill-comply; whole-config GC → config-gc.
learn-eval
Extract a reusable pattern from the current session, judge it Save / Improve then Save / Absorb / Drop against a grounding checklist, and route every Save to a destination something actually reaches — absorbed into an existing skill / rule / doc section, or promoted to a real skill via skill-creator. Use when the user says 「今回の学びを残して」「learn-eval して」 or /learn-eval. There is no notes parking lot: if nothing would route to it, the verdict is Drop. NOT for mining past sessions (session-judgment-mining), auditing skills (skill-stocktake), or distilling rules (rules-distill).
rules-distill
Scan installed skills to extract principles that belong in the always-loaded rules layer (environment-specific facts, wiring, and traps — not general principles the substrate already applies) and distill them into rules — append to, revise, or create rule files. Use when the user says "distill rules", "/rules-distill", "promote patterns to rules", "what principles should become rules", after installing new skills, or when a skill-stocktake surfaces recurring patterns. NOT for auditing skill quality (that is skill-stocktake) and NOT for editing a single skill (that is skill-creator).
rules-stocktake
Audit ~/.claude/rules (always-loaded behavioral rules) for residency cost, staleness, redundancy, broken See-skill pointers, and substrate absorption, assigning Keep/Improve/Update/Merge/Demote-to-skill/Dissolve/Retire verdicts. Use when the user says "audit my rules", "rules stocktake", "which rules should be demoted or dissolved", 「rules が肥大化してきた」「ルールを棚卸しして」, or when the model generation changed and over-constraints written for the previous one may now be net-negative (「新しいモデルに合わせて rules を見直したい」「rightsize したい」). NOT for — skill quality → skill-stocktake; promoting skill patterns INTO rules → rules-distill (this is its inverse); runtime compliance → skill-comply; whole-config GC → config-gc.
skill-comply
Visualize whether skills, rules, and agent definitions are actually followed — auto-generates scenarios at 3 prompt strictness levels, runs agents, classifies behavioral sequences, and reports compliance rates with full tool call timelines
skill-stocktake
Audit installed Claude skills for quality and surface Keep/Improve/Update/Retire/Merge verdicts. Use when the user says "audit my skills", "stocktake", "review my skills", "which skills should I retire or merge", "do a quality pass over my skills", or "/skill-stocktake". NOT for creating or improving a single skill (that is skill-creator) and NOT for whole-config GC across hooks/permissions/MCP (that is config-gc).
llm-as-judge
Design pattern for LLM-as-judge evaluators — binary checks as evidence, one named holistic verdict, no score aggregation. Use when designing or reviewing any LLM-based quality gate, evaluator, judge prompt, or verdict schema; when a judge's rubric scores fluctuate between runs; when you catch yourself asking an LLM for a 1-5 score, averaging check results, or thresholding a satisfaction ratio. NOT for choosing whether a task needs deterministic or semantic processing, and NOT for the architecture-level judge+enforce state-mutation split.
skill-health
Scan the skill library for structural technical debt — dangling references where a SKILL.md names a script, bash file, agent, or sibling skill that does not exist on disk ("missing artifacts" debt), plus external skills whose directory is a symlink out of the skills root, where a local fix would be overwritten by the owning package manager. Use when the user says "scan skills for debt", "check for dangling/broken references in my skills", "skill health check", "do referenced scripts/agents still exist", "which skills are not really mine to edit", or "/skill-health". NOT for holistic skill quality verdicts (that is skill-stocktake), NOT for config GC over hooks/permissions/MCP (that is config-gc), and NOT for security scanning (that is the /claude-security plugin).
paper-ecosystem
Academic paper / preprint / position paper の執筆・レビューエコシステムの orchestrator。SSRN / arXiv / Zenodo / journal venue 向けの **harness-neutral な学術コンテンツ** を書く / レビューするときに使う。paper-writing / paper-reviewer / source-fidelity-checker / vocabulary-consistency-checker / clarity-reviewer / citation-formatter の役割境界と使い分け、Source Fidelity Rules、Vocabulary Consistency Rules、Academic Voice Rules、Reader Clarity Rules、Citation Format Rules を正本として保持する。人間向け blog / essay には `writing-ecosystem`、AI 向け doc には `llms-txt-writer` を使う。
writing-ecosystem
人間向け記事・エッセイ・ブログポスト・ニュースレターの唯一の執筆 orchestrator。project の publication channel contract を読み、中心命題 1 つの editorial brief、因果線、証拠の選択と除外、構成、執筆、title-reviewer、review panel、quality-gate、著者 GO までを統括する。Use when — 「この記事を書いて」「このテーマでエッセイにして」「原稿の論点を一つに絞って構造改稿して」のような新規執筆・全体改稿・全文の別 channel 展開。NOT for — 一文や段落だけの翻訳（→ prose-translation）、title だけ（→ headline-craft / title-reviewer）、SNS 下書き（→ x-draft）、公開 thread 返信（→ public-comment）、AI 向け docs、README、paper、媒体固有の公開操作。
paper-writing
Academic paper / preprint / position paper の draft skill。Title + outline + section drafting + abstract + references の手順正本。Primary source への直接 access を強制し、claim と cite の 1:1 mapping を author に握らせる。SSRN / arXiv / Zenodo / journal venue 向けに使う。Voice 規約・Source Fidelity 規約・Citation 規約は orchestrator skill `paper-ecosystem` を参照する。
search-first
Search npm / PyPI / MCP / GitHub / existing skills before writing code for any new feature, integration, library selection, or utility. Use whenever the user says 'add X functionality', 'implement Y', 'set up Z', 'integrate W', asks 'what library should I use for...', 'is there a package/client/MCP for...', proposes a specific tool while open to alternatives, or hits the planning.md wiring that routes dependency additions and self-built utilities here first. Especially when the task adds a dependency, picks between tools, or builds a utility that likely already exists (parser, checker, converter, linter, CI step, E2E framework, payment / auth / API client). DO NOT use for bug fixes, refactoring, config value edits, file summarization, or throwaway scripts whose approach the user already fully specified.
search-first
Search npm / PyPI / MCP / GitHub / existing skills before writing code for any new feature, integration, library selection, or utility. Use whenever the user says 'add X functionality', 'implement Y', 'set up Z', 'integrate W', asks 'what library should I use for...', 'is there a package/client/MCP for...', proposes a specific tool while open to alternatives, or hits the planning.md wiring that routes dependency additions and self-built utilities here first. Especially when the task adds a dependency, picks between tools, or builds a utility that likely already exists (parser, checker, converter, linter, CI step, E2E framework, payment / auth / API client). DO NOT use for bug fixes, refactoring, config value edits, file summarization, or throwaway scripts whose approach the user already fully specified.
citation-sync
研究 repo の引用 3 層 (docs 実引用 → .zenodo.json references → graph.jsonld ExternalReference) を監査し、下層から順に同期する orchestrator。旧層 4 (Wikidata P2860) は 2026-07 の governance revocation により恒久 retire (authorship-strategy ADR-0021) — 同期対象にしない。Use when the user says 「参考文献が少ない/足りない」「引用がずれてる」「citation を同期して」「引用文献を graph に反映して」, when new external papers were cited in ADR/glossary/empirical docs, or before a release of a DOI-registered repo. 単層の実装は release-doi (.zenodo.json) / jsonld-knowledge-graph (graph) に defer し、本 skill は層間の divergence 検出・curation 基準・同期順序だけを持つ。NOT for: 論文 (paper item) 自体の reference list 整備 (paper-deposit が担当)、引用を含まない repo。
when-code-when-llm
Decision framework for choosing between deterministic code (regex, keyword match, AST parse, schema validation) and LLM-based processing (classification, semantic similarity, judge) for a single task. Use when you catch yourself writing a regex for a task that keeps producing false positives or negatives, or when you are about to call an LLM for something a three-line code check would handle. Covers the structural-vs-semantic axis, the false-positive test, worked examples of both directions, and the enumerate/decide split for tasks where detection is structural but resolution needs judgment.
measurement-discipline
測定に基づく主張・閾値・ガード・実験結果を設計または評価するときの規律。Use when the user says 「この実験結果で判断していい？」「閾値を決めたい」「ガード/検査を足したい」「1 回通ったから大丈夫」, when a design places a numeric threshold or a suspicion flag, or when a claim rests on measured data. NOT for — 計器（read-only 分布・読み値）そのものの設計（CA repo の skill read-only-instruments が正本）、LLM 判定器の設計（llm-as-judge）、ループ構造の妥当性（loop-design-check）。
repair-discipline
バグ修正・残課題・schema/storage 変更に着手するときの規律。Use when the user says 「このバグ直して」「残課題をやって」「この schema を変えたい」, when picking up a stale task file, or when a fix touches storage formats or shared gates. NOT for — chain の種別とレビュー条件の判定（implementation-chain）、TDD の手順（tdd）、Python 固有の mock / fixture の罠（python-patterns）、台帳全体の棚卸し（task-stocktake — 本 skill は 1 件着手時の照合のみ）。
prompt-perturb
同じ手筋・いつもの skill ばかりで進む作業に、文脈を持たない検索者（prompt-forager agent）が外部の創造技法カタログ・プロンプト集から異角度プロンプトを取ってきて注入する多様性注入 skill。ブレインストーミングしたい、いつもと違う角度のアイデアが欲しい、発想がマンネリ・局所解に嵌っている、同じようなアプローチばかり出てくる、と感じたとき、または /prompt-perturb <目的の一行> で使う。正解・最適解の探索には使わない（それは search-first / scout）、プロンプトの品質改善にも使わない（それは prompt-writer agent）。
context-sync
Audit and fix project documentation — detect role overlaps between context files (CLAUDE.md, CODEMAPS, ADR, README), migrate misplaced content, check freshness against code, and create missing docs. One command to keep all project context healthy.
learn-eval
Extract a reusable pattern from the current session, judge it Save / Improve then Save / Absorb / Drop against a grounding checklist, and route every Save to a destination something actually reaches — absorbed into an existing skill / rule / doc section, or promoted to a real skill via skill-creator. Use when the user says 「今回の学びを残して」「learn-eval して」 or /learn-eval. There is no notes parking lot: if nothing would route to it, the verdict is Drop. NOT for mining past sessions (session-judgment-mining), auditing skills (skill-stocktake), or distilling rules (rules-distill).
repo-asset-stocktake
Audit a project repo's non-code assets — tool configs, CI/GitHub workflows, runbooks, other docs — for assets whose consumer has vanished, and assign Keep/Update/Retire/Merge verdicts. Use when the user says "audit my repo assets", "which configs/workflows/runbooks are dead", "repo asset stocktake", 「非コード資産を棚卸しして」「使われてない設定/workflow/runbook を洗い出して」. NOT for — dead code → refactor-clean; doc-role overlap across CLAUDE.md/CODEMAPS/ADR/README → context-sync; ~/.claude config GC → config-gc; skills or rules → skill-stocktake / rules-stocktake.
rules-distill
Scan installed skills to extract principles that belong in the always-loaded rules layer (environment-specific facts, wiring, and traps — not general principles the substrate already applies) and distill them into rules — append to, revise, or create rule files. Use when the user says "distill rules", "/rules-distill", "promote patterns to rules", "what principles should become rules", after installing new skills, or when a skill-stocktake surfaces recurring patterns. NOT for auditing skill quality (that is skill-stocktake) and NOT for editing a single skill (that is skill-creator).
rules-stocktake
Audit ~/.claude/rules (always-loaded behavioral rules) for residency cost, staleness, redundancy, broken See-skill pointers, and substrate absorption, assigning Keep/Improve/Update/Merge/Demote-to-skill/Dissolve/Retire verdicts. Use when the user says "audit my rules", "rules stocktake", "which rules should be demoted or dissolved", 「rules が肥大化してきた」「ルールを棚卸しして」, or when the model generation changed and over-constraints written for the previous one may now be net-negative (「新しいモデルに合わせて rules を見直したい」「rightsize したい」). NOT for — skill quality → skill-stocktake; promoting skill patterns INTO rules → rules-distill (this is its inverse); runtime compliance → skill-comply; whole-config GC → config-gc.
search-first
Search npm / PyPI / MCP / GitHub / existing skills before writing code for any new feature, integration, library selection, or utility. Use whenever the user says 'add X functionality', 'implement Y', 'set up Z', 'integrate W', asks 'what library should I use for...', 'is there a package/client/MCP for...', proposes a specific tool while open to alternatives, or hits the planning.md wiring that routes dependency additions and self-built utilities here first. Especially when the task adds a dependency, picks between tools, or builds a utility that likely already exists (parser, checker, converter, linter, CI step, E2E framework, payment / auth / API client). DO NOT use for bug fixes, refactoring, config value edits, file summarization, or throwaway scripts whose approach the user already fully specified.
skill-comply
Visualize whether skills, rules, and agent definitions are actually followed — auto-generates scenarios at 3 prompt strictness levels, runs agents, classifies behavioral sequences, and reports compliance rates with full tool call timelines
skill-health
Scan the skill library for structural technical debt — dangling references where a SKILL.md names a script, bash file, agent, or sibling skill that does not exist on disk ("missing artifacts" debt), plus external skills whose directory is a symlink out of the skills root, where a local fix would be overwritten by the owning package manager. Use when the user says "scan skills for debt", "check for dangling/broken references in my skills", "skill health check", "do referenced scripts/agents still exist", "which skills are not really mine to edit", or "/skill-health". NOT for holistic skill quality verdicts (that is skill-stocktake), NOT for config GC over hooks/permissions/MCP (that is config-gc), and NOT for security scanning (that is the /claude-security plugin).
skill-stocktake
Audit installed Claude skills for quality and surface Keep/Improve/Update/Retire/Merge verdicts. Use when the user says "audit my skills", "stocktake", "review my skills", "which skills should I retire or merge", "do a quality pass over my skills", or "/skill-stocktake". NOT for creating or improving a single skill (that is skill-creator) and NOT for whole-config GC across hooks/permissions/MCP (that is config-gc).
rules-stocktake
Audit ~/.claude/rules (always-loaded behavioral rules) for residency cost, staleness, redundancy, broken See-skill pointers, and substrate absorption, assigning Keep/Improve/Update/Merge/Demote-to-skill/Dissolve/Retire verdicts. Use when the user says "audit my rules", "rules stocktake", "which rules should be demoted or dissolved", 「rules が肥大化してきた」「ルールを棚卸しして」, or when the model generation changed and over-constraints written for the previous one may now be net-negative (「新しいモデルに合わせて rules を見直したい」「rightsize したい」). NOT for — skill quality → skill-stocktake; promoting skill patterns INTO rules → rules-distill (this is its inverse); runtime compliance → skill-comply; whole-config GC → config-gc.
authorship-strategy
maker / 実践者が AI 時代の著者戦略を実地で探るための判断フレームワーク。あなた自身の DOI-registered idea-rescue repo 群（著者自身の例は AKC, Contemplative Agent）で適用する。core principle は「AI 時代のオーセンティシティ inversion」— 3 軸（scarcity → diffusion / exclusivity → derivation / enclosure → openness）の反転。主 audience は層別（ADR-0022 二層会計）— doctrine 層 = LLM-mediated channels（LLM 直接 + LLM 経由で情報を得る人間）、essay 層 = 同時代の human reader（同時代 KPI の観測+操舵可）。creative reuse > training > investigation の preference 階層。4 層 framework (authenticity → diffusion → idea/scaffold 判別 → tactics) で判断軸を提供。マネタイズ禁止、tool-agnostic、any-usage 容認、permissive license 原則
release-doi
DOI-registered research repo (Zenodo) のリリース手順。CODEMAPS / README 多言語 / CHANGELOG / CITATION.cff / pyproject.toml / llms.txt / glossary を整合させてから tag push、Zenodo 自動採番後に新 DOI を反映し、Software Heritage archive + SWHID 記録 (intrinsic identifier 層) まで行う 5 phase + post-release ワークフロー。AKC / AAP / contemplative-agent など shimo4228 系の研究 repo で再利用する。
readme-writer
README やプロジェクトのトップページ（repo を開いた人が最初に見る入口）を書く・直すときに使う。こんな時に呼ぶ — README が長い／継ぎ足しで文脈が重くなり初見で分からない、開いて数十秒で「何のプロジェクトで自分向けか」が伝わる入口にしたい、長い rationale・ADR 参照・内部史を docs/ に逃がしたい、研究・DOI repo の README を引用付きで読める長さにまとめたい、GitHub の About（description / topics）が README と食い違っている。証拠スクリプト（readme_evidence.py）+ fresh context の判定器（readme-judge）+ 上限 2 ラウンドの改稿ループ + review panel + 著者通読 GO で回す。短く・走査しやすくしつつ、LLM が README 一枚で要点を復元できる情報フロアは残す。CLI でも研究 repo でも、日本語でも英語でも、新規作成でも既存改善でも対象。AI 専用ドキュメント（llms.txt 等）は → llms-txt-writer、記事・エッセイは → writing-ecosystem。
context-sync
Audit and fix project documentation — detect role overlaps between context files (CLAUDE.md, CODEMAPS, ADR, README), migrate misplaced content, check freshness against code, and create missing docs. One command to keep all project context healthy.
skill-comply
Visualize whether skills, rules, and agent definitions are actually followed — auto-generates scenarios at 3 prompt strictness levels, runs agents, classifies behavioral sequences, and reports compliance rates with full tool call timelines
rules-distill
Scan installed skills to extract principles that belong in the always-loaded rules layer (environment-specific facts, wiring, and traps — not general principles the substrate already applies) and distill them into rules — append to, revise, or create rule files. Use when the user says "distill rules", "/rules-distill", "promote patterns to rules", "what principles should become rules", after installing new skills, or when a skill-stocktake surfaces recurring patterns. NOT for auditing skill quality (that is skill-stocktake) and NOT for editing a single skill (that is skill-creator).
learn-eval
Extract a reusable pattern from the current session, judge it Save / Improve then Save / Absorb / Drop against a grounding checklist, and route every Save to a destination something actually reaches — absorbed into an existing skill / rule / doc section, or promoted to a real skill via skill-creator. Use when the user says 「今回の学びを残して」「learn-eval して」 or /learn-eval. There is no notes parking lot: if nothing would route to it, the verdict is Drop. NOT for mining past sessions (session-judgment-mining), auditing skills (skill-stocktake), or distilling rules (rules-distill).
skill-health
Scan the skill library for structural technical debt — dangling references where a SKILL.md names a script, bash file, agent, or sibling skill that does not exist on disk ("missing artifacts" debt), plus external skills whose directory is a symlink out of the skills root, where a local fix would be overwritten by the owning package manager. Use when the user says "scan skills for debt", "check for dangling/broken references in my skills", "skill health check", "do referenced scripts/agents still exist", "which skills are not really mine to edit", or "/skill-health". NOT for holistic skill quality verdicts (that is skill-stocktake), NOT for config GC over hooks/permissions/MCP (that is config-gc), and NOT for security scanning (that is the /claude-security plugin).
paper-ecosystem
Academic paper / preprint / position paper の執筆・レビューエコシステムの orchestrator。SSRN / arXiv / Zenodo / journal venue 向けの **harness-neutral な学術コンテンツ** を書く / レビューするときに使う。paper-writing / paper-reviewer / source-fidelity-checker / vocabulary-consistency-checker / clarity-reviewer / citation-formatter の役割境界と使い分け、Source Fidelity Rules、Vocabulary Consistency Rules、Academic Voice Rules、Reader Clarity Rules、Citation Format Rules を正本として保持する。人間向け blog / essay には `writing-ecosystem`、AI 向け doc には `llms-txt-writer` を使う。
paper-writing
Academic paper / preprint / position paper の draft skill。Title + outline + section drafting + abstract + references の手順正本。Primary source への直接 access を強制し、claim と cite の 1:1 mapping を author に握らせる。SSRN / arXiv / Zenodo / journal venue 向けに使う。Voice 規約・Source Fidelity 規約・Citation 規約は orchestrator skill `paper-ecosystem` を参照する。
repo-asset-stocktake
Audit a project repo's non-code assets — tool configs, CI/GitHub workflows, runbooks, other docs — for assets whose consumer has vanished, and assign Keep/Update/Retire/Merge verdicts. Use when the user says "audit my repo assets", "which configs/workflows/runbooks are dead", "repo asset stocktake", 「非コード資産を棚卸しして」「使われてない設定/workflow/runbook を洗い出して」. NOT for — dead code → refactor-clean; doc-role overlap across CLAUDE.md/CODEMAPS/ADR/README → context-sync; ~/.claude config GC → config-gc; skills or rules → skill-stocktake / rules-stocktake.
when-code-when-llm
Decision framework for choosing between deterministic code (regex, keyword match, AST parse, schema validation) and LLM-based processing (classification, semantic similarity, judge) for a single task. Use when you catch yourself writing a regex for a task that keeps producing false positives or negatives, or when you are about to call an LLM for something a three-line code check would handle. Covers the structural-vs-semantic axis, the false-positive test, worked examples of both directions, and the enumerate/decide split for tasks where detection is structural but resolution needs judgment.
citation-sync
研究 repo の引用 3 層 (docs 実引用 → .zenodo.json references → graph.jsonld ExternalReference) を監査し、下層から順に同期する orchestrator。旧層 4 (Wikidata P2860) は 2026-07 の governance revocation により恒久 retire (authorship-strategy ADR-0021) — 同期対象にしない。Use when the user says 「参考文献が少ない/足りない」「引用がずれてる」「citation を同期して」「引用文献を graph に反映して」, when new external papers were cited in ADR/glossary/empirical docs, or before a release of a DOI-registered repo. 単層の実装は release-doi (.zenodo.json) / jsonld-knowledge-graph (graph) に defer し、本 skill は層間の divergence 検出・curation 基準・同期順序だけを持つ。NOT for: 論文 (paper item) 自体の reference list 整備 (paper-deposit が担当)、引用を含まない repo。
adr-writer
Record a design decision as an Architecture Decision Record (ADR) in the project's `docs/adr/` directory. Use this skill whenever the user says "let's ADR this", "record this decision", "write an ADR for X", or when context-sync Phase 3 needs to extract a buried decision. The skill resolves the target ADR directory from cwd, picks the next sequence number with no collision, delegates 7-section body generation (incl. `Review-when` expiry conditions) to the adr-writer agent, and updates the ADR index. Works across any repo — auto-detects or creates `docs/adr/` from the repo root.
ai-native-preprint-submission
AI-native 出版プラットフォーム (aiXiv / AiraXiv 等、AI 審査・agent 読者を持つ preprint archive) へ、deposit 済み論文を投稿するワークフロー。経路は 2 つ — (A) Web UI + browser automation (初回投稿の default。Claude がフォーム入力を代行し、アカウント作成・ファイル選択・送信クリックは著者が行う人間ゲート分担)、(B) 著者発行の API key による MCP 投稿 (2 本目以降・著者が明示委任した場合)。Use when the user says 「aiXiv/AiraXiv に投稿して」「AI-native プラットフォームに論文を出して」「API キーで残りを投稿して」or invokes /ai-native-preprint-submission. NOT for: Zenodo/SSRN への正本 deposit (それは paper-deposit)、arXiv 投稿 (endorsement 制で別物)。
cited-source-mirror-verification
Verify a cited empirical/numeric claim against an open mirror (aiXiv, OpenAlex, Semantic Scholar, the author's own copy) before citing it. Use whenever you are about to cite a statistic, effect size, odds ratio, percentage, p-value, or empirical result you have NOT read in the primary — because the primary is access-blocked (SSRN / publisher 403, paywalled PDF), or because the claim arrived via an LLM research digest, daily-research note, or WebSearch summary (which routinely reframe a methodological artifact or a limitation as a headline finding). Especially before committing citations to a durable artifact — paper, knowledge graph, .zenodo.json, ADR, DOI repo.
collect-context
記事・エッセイを書く前の素材収集。セッション内外のコンテキストを集め、全項目ソース付きの証拠台帳（evidence dossier）— Claims Register・一次/⚠未検証の tier・セッションログ索引 — を生成する。Use when — 「素材を集めて」「証拠台帳を作って」「この作業の記事コンテキストをまとめて」、執筆に着手する前、fact-checker に渡す主張リストが要るとき。NOT for — 過去ログから記事の問いを発見（→ session-theme-mining）、テーマ・構成・タイトル等の編集判断（受け側 repo の責務。正本は受け側 repo の rules のチャンネル表、執筆 orchestrator は writing-ecosystem — zenn-content 常駐）、事実の検証そのもの（→ fact-checker agent）。収集者は推薦・提案・方向性メモを出力に書かない
en-to-ja-translation
英語→日本語の voice 保持翻訳スキル。エッセイ・研究ドキュメント・README・ADR 等の人間向け prose を、著者の声・register・発見調を保ったまま自然な日本語にする。逐語訳でも MT でもなく、term-lock（訳す-by-default／英語保持は明示例外）+ 脱翻訳調 pass +（EN→JA→EN の）back-translation QA で「英語のまま読みづらい」「専門用語が英語のまま」を潰す。英語記事を日本語にする EN→JA 翻訳タスクで使う。AI 向け doc は llms-txt-writer、学術 citation format は citation-formatter、日本語 AI-slop / Voice 規約は writing-ecosystem に defer。
harness-sync
ローカル harness (~/.claude) の origin-filtered コンポーネントを公開 repo (claude-harness 集約 repo + 単独 skill repo 群) へ一方向同期する。Use when the user says 「ハーネスを公開 repo に同期して」「claude-harness を更新して」「スキルを公開して」「skill repo を同期して」 or invokes /harness-sync. 収集 → secret scan → subtree 置換は決定論的 script が行い、diff レビュー・README/llms.txt の整合・コミットは会話で行う。NOT for: 公開 repo から ~/.claude への逆方向取り込み、ECC 等外部 origin の公開判断、汎用化 fork を持つ curated skill repo (script 同期は汎用化を壊すため手動 curation)。
headline-craft
「開かせる一行」を作る候補生成スキル。記事タイトル・README tagline・subtitle・SNS告知文を、具体性・ベネフィット・誠実な好奇心ギャップ・検索/フィードの技法から生成する。Use when the user asks for タイトル案・キャッチコピー・タグライン・見出し候補、or a frozen draft needs title candidates. NOT for — 公開記事候補の点検（→ title-reviewer）、煽り・クリックベイト、topics / emoji、platform文字数の定義。
hf-sync
Hugging Face Datasets mirror sync for graph.jsonld-bearing research repos. `<Owner/dataset>` を引数に取り、cwd の graph.jsonld を flatten して graph.jsonl と一緒に HF dataset へ upload する。`release-doi` Phase 4 末尾の `gh release create` 後、または ad-hoc resync で起動する（tag push は webhook を発火させないので契機にしない）。Local の `hf auth login` token を使うので GitHub Actions / CI auth は不要。
implementation-chain
実装に着手する前に task 種別（feat / fix / refactor / chore / prototype / writing）を判定し、その種別に対応する agent chain（Plan → Phase 0 → TDD → Review 群 → Doc Sync → Verify）を plan に front-load するための判断表。Use when starting to implement a feature, fix a bug, refactor, or write a document and you need to decide which reviewers and gates apply — 「これから実装する」「chain を組む」「どのレビューを回すべきか」。writing 種別の orchestrator skill へのルーティング表と早期停止条件もここが正本。NOT for — chain 内の各ステップの実装詳細（それは search-first / tdd / codex-review / writing-ecosystem 等の各 skill）、既に chain が確定した後の実行。
ja-to-en-translation
日本語⇄英語の voice 保持翻訳スキル（**両方向**）。エッセイ・研究ドキュメント・README・ADR 等の人間向け prose を、著者の声・register・発見調を保ったまま自然な訳文にする。逐語訳でも MT でもなく、term-lock + 2-pass（訳→自己添削）+ back-translation QA で品質を担保する。JA→EN は英語 AI-slop の自己添削、EN→JA は訳す-by-default の term policy と脱翻訳調 pass（英語語順残存・冗長受動態・カタカナ乱用・「の」連鎖・直訳 idiom・接続の機械訳）を追加で適用する。日本語記事を英語にするとき、英語記事・EN 正本 README を日本語にするときに使う。AI 向け doc は llms-txt-writer、学術 citation format は citation-formatter、AI-slop / Voice / Title 規約と出典編入は writing-ecosystem、出力先の語尾はチャンネル表に defer。
paper-deposit
Deposit a finished, reviewed academic paper / position paper / preprint to a DOI registry (Zenodo) as a standalone record, optionally cross-post it to SSRN for extra reach, generate the PDF from Markdown, and cross-link the minted DOI back into a research-program repository (.zenodo.json / README / graph.jsonld / llms.txt). Use this whenever the user wants to "ship", "deposit", "publish", "register a DOI for", "put on Zenodo", or "submit to SSRN" a paper that is already written and review-passed — even if they only say "let's get this paper out". This runs AFTER paper-ecosystem review passes. It is NOT for releasing a whole repository (that is release-doi) and NOT for drafting or reviewing the paper (that is paper-writing / paper-ecosystem).
public-comment
公開の技術ピア議論への返信コメントを書く skill。GitHub discussions / issues / PR コメント・Hugging Face discussions・技術フォーラムなど、初対面の技術者が読む公開スレッドへの返信を作成・改稿・投稿するときに使う。Use when the user says 「GitHub のコメントに返信して」「discussion にコメント書いて」「issue に返信」「このスレッドに参加したい」 or asks to draft/revise/post a reply in a public technical thread. 2026 年の OSS は AI slop 危機下にあり、tell（em-dash・対比構文・triad 等）は「未レビューの生ダンプ」のシグナルとして読まれる — 本 skill は content ownership・スレッド接地・脱 tell 改稿・日本語訳併記の人間 gate までを扱う。NOT for: 記事・エッセイ（→ writing-ecosystem）、AI 向け doc（→ llms-txt-writer）、既知の相手との Slack/Discord 会話、コードの PR 本体。
task-stocktake
Consolidate a repo's pending-task tracking into its single task ledger (default .notes/TASKS.md) — bootstrap it if missing, sweep handoff / audit / remaining-issues files and auto-memory for stray task lines, verify pending entries against git log and actual code, archive detail files of completed tasks. Use when the user says 「残タスクを棚卸しして」「タスク台帳を作って/整理して」「残っているタスクは？」, "task stocktake", when task lines are scattered across notes files, or when a repo's ledger may be stale. NOT for — skills → skill-stocktake; rules → rules-stocktake; repo non-code assets → repo-asset-stocktake; in-session todos → harness task tools.
wiki-query
Obsidian Vault の LLM wiki (wiki/concept/) に問い合わせ、[[ ]] 出典付きで合成回答する query。vault セッションだけでなく研究 repo (AKC / AAP / contemplative / authorship 等) のセッションからも呼べる。複数ページを横断して合成した良回答は書き戻しプロトコルで wiki/concept/ に filing する（Karpathy 原典 parity、read-write）。Use when the user invokes /wiki-query <問い>, asks 「wiki に聞いて」「wiki ではどうなってる？」, or when working in a research repo and a synthesized understanding of past daily-research notes would answer the question faster than grep. NOT for source の新規取り込み・wiki 全体の健全性チェック (それらは vault セッション専用の /ingest・/lint-wiki)。
x-draft
daily-research のリサーチレポートから X (Twitter) 投稿用の日本語長文下書きをオンデマンド生成する対話 skill。Use when the user says 「X に投稿したい」「X 用の下書き作って」「これポストしたい」「ツイートにして」, invokes /x-draft, or mentions turning a daily-research report into a social media post — even if they just say 「今朝の akc のやつ X に流したい」 like shorthand. pull 型: 通知もノルマもなく、ユーザーがレポートを読んで投稿したいと思ったときだけ呼ぶ。軽量事故ゲート (一次ソース再確認 + 3 日陳腐化チェック) と脱 AI-tell 改稿を通した、コピペ 1 回で投稿できる長文 1 ポストを出力する。NOT for: 記事・エッセイ執筆（→ writing-ecosystem）、公開スレッドへの返信（→ public-comment）、タイトル・一行コピーだけの生成（→ headline-craft）、X への自動投稿（本 skill ��下書きまで — 投稿は人間が行う）。
prose-translation
日本語⇄英語の voice 保持翻訳スキル（**両方向**）。エッセイ・記事・README・ADR 等の人間向け prose を、出力先の publication channel contract が宣言する register と原文の確度を保って自然に訳す。逐語訳でも MT でもなく、term-lock + 2-pass（訳→自己添削）+ back-translation QA で品質を担保する。JA→EN は英語 AI-slop の自己添削、EN→JA は訳す-by-default の term policy と脱翻訳調 passを追加する。AI 向け doc は llms-txt-writer、学術 citation format は citation-formatter、shared craft は writing-ecosystemへ defer。
rfc-writer
公開 rfcs/ 台帳へ 1 エントリを起票する手順と規約の唯一の正本（足切り → 採番 → 様式 → 公開規約 → spawn 接続 → index 行）。Use when the user says 「これ起票して」「RFC にしておいて」「提案を台帳に載せて」, when harness-boundary の Defer や task-triage の起票提案が承認されたとき, or /rfc-writer. 各 repo の rfcs/README.md は薄いポインタ + index のみで、規約本文はここ以外に書かない。NOT for — 状態語彙の定義と台帳の棚卸し（→ task-stocktake が正本）、起票するかの足切り判定のうち review 指摘の loop-breaking + producer 規律（→ task-stocktake「レビュー指摘の起票規律」）、open タスクの判定・dispatch・検収（→ task-triage）、決定の記録（→ adr-writer。rfcs は提案・未決、ADR は決定）、単一表 .notes/TASKS.md への 1 行起票（→ rule task-tracking の形のまま）。
review-to-lint
既存 reviewer（agent / review skill）のチェックリスト、または過去セッションの reviewer 履歴から機械判定可能な項目を決定論 script に抽出し、reviewer を意味的チェック専任に薄化する手順。著者が「このレビュアーを lint 化して」「レビューを lint に吸収して」「機械チェックを script に降ろして」「履歴から lint 化できるものを探して」と言ったとき、または reviewer の指摘に機械的項目の反復が目立つときに /review-to-lint で使う。NOT for — 新規 reviewer の作成（→ skill-creator）、意味的基準そのものの変更（各 reviewer の正本）、judge の設計（→ llm-as-judge）、既に evidence script を持つ reviewer の再抽出（readme-writer / adr-writer は実施済み）。
session-theme-mining
過去の Claude Code / Codex セッションを横断し、記事になりうる未解決の問いを 0〜3 件の同格な候補カードとして発見する。Use when — 「過去セッションから記事テーマを探して」「まだ書いていない問いを発掘して」「セッション履歴から collect-context の入口を作って」。NOT for — 選択済みテーマの証拠収集（→ collect-context）、ユーザーの価値観を skill / rule に昇格（→ session-judgment-mining）、候補の採点・順位付け・タイトル作成。
writing-ecosystem
人間向け記事・エッセイ・ブログポスト・ニュースレターの唯一の執筆 orchestrator。project の publication channel contract を読み、中心命題 1 つの editorial brief、因果線、証拠の選択と除外、構成、執筆、review panel、著者の内容 GO、title-reviewer、quality-gate までを統括する。Use when — 「この記事を書いて」「このテーマでエッセイにして」「原稿の論点を一つに絞って構造改稿して」のような新規執筆・全体改稿・全文の別 channel 展開。NOT for — 一文や段落だけの翻訳（→ prose-translation）、title だけ（→ headline-craft / title-reviewer）、SNS 下書き（→ x-draft）、公開 thread 返信（→ public-comment）、AI 向け docs、README、paper、媒体固有の公開操作。
codex-review
Cross-model second opinion from the OpenAI Codex CLI (a different model family), read-only, in two seams — (1) code review of the current diff; (2) plan-stage premise challenge of a design packet (refute / missing / alternative, never a design). Opt-in only — use ONLY when the user explicitly asks with "codex review", "cross-model review", "second opinion on this diff", "別モデルでレビュー", "プランを Codex に反証させて", "前提を別モデルで叩いて", or invokes /codex-review or /codex-review --plan <file>. NOT a default step of the implementation chain (自発発火しない), NOT for letting Codex write code or design (read-only, divergence only), and NOT a replacement for the in-Claude reviewers.
spawn-session
新しい detached な Claude Code Remote Control セッションを Herdr 内に起動し、Claude モバイルアプリのセッション一覧に出す。生きている任意のセッションから（多くは iPhone の Remote Control 越しに）呼んで、別プロジェクトの新規セッションを Mac に触れず立ち上げる。Use when the user says 「新しいセッション立てて」「AAP のセッション開いて／立ち上げて」「contemplative のセッション作って」「spawn a (new) session」「launch a remote control session」「start a session for X」, or invokes `/spawn-session [project]`. NOT for: 既存会話の resume（`--continue`/`--resume`）、同一セッション内の文脈リセット（`/clear`）、現セッションの model 切替。
session-theme-mining
過去の Claude Code / Codex セッションを横断し、記事になりうる未解決の問いを 0〜3 件の同格な候補カードとして発見する。Use when — 「過去セッションから記事テーマを探して」「まだ書いていない問いを発掘して」「セッション履歴から collect-context の入口を作って」。NOT for — 選択済みテーマの証拠収集（→ collect-context）、ユーザーの価値観を skill / rule に昇格（→ session-judgment-mining）、候補の採点・順位付け・タイトル作成。
herdr-delegate
Herdr の pane に別プロセスの CLI コーディングエージェント（Codex 等）を立てて、実装タスクを丸ごと委譲するワークフロー。Use when the user says 「Codex にやらせて」「E2E は Codex に委譲」「Herdr でセッション立ててタスク投げて」 or explicitly asks to delegate implementation work to another CLI agent. 前提ゲート: HERDR_ENV=1 かつユーザーの明示指示（agents.md の委譲ゲートと同一 — 有益そうというだけで自発起動しない）。herdr CLI の一般操作は skill: herdr が正本で、本 skill は委譲に特化した手順・監視レシピ・検収規律のみを持つ。NOT for — read-only の cross-model レビュー（→ codex-review）、Claude 内サブエージェントへの並列化（→ Agent tool）。
substack-publishing
完成・レビュー済みの human essay を Substack に公開し、LLM 発見のために corpus へミラーするワークフロー。Substack が raw Markdown 非対応なための MD→HTML rich-text paste、Title/Subtitle/body のフィールド分け、タグ戦略（archive 用 ≠ 拡散用）、カバー画像プロンプトの作り方、Claude in Chrome によるエディタ自動操作（OS クリップボードへの HTML flavor 直接セット + cmd+V）、公開後の content repo `substack/` フォルダへのミラー + research repo からの cross-link、公開後の配信ファネル運用（Notes 3 型・organic recommendations・welcome email・ケイデンス）を扱う。Voice / AI-slop / Title / 出典は writing-ecosystem、翻訳は ja-to-en-translation に defer。essay を Substack に出すとき・Substack の配信運用を考えるときに使う。
signal-first-research
Design guide for a research intake filter that admits only information likely to change your next action. Use when you are about to build a recurring research workflow (daily digest, news stream, topic monitor, literature feed) and the default question "what should I read?" keeps producing too much. Covers the define-signal-first rule, three worked examples, and a diagnostic checklist for telling a healthy filter from a lazy one.
llm-agent-security-principles
Structural security design principles for building LLM agents, autonomous systems, and self-improving harnesses. Use when designing a new agent, reviewing agent architecture, adding a new capability, or auditing an existing agent for prompt injection resistance. Covers three load-bearing principles — Security by Absence, Untrusted Content Boundary, One External Adapter — grounded in the prohibition-strength hierarchy (absence > scaffolding enforcement > untrusted boundary), plus concrete defense patterns for HTTP, credentials, and LLM hosts.
code-and-llm-collaboration
Design patterns for layering deterministic code and LLM calls in a single pipeline or agent. Use when designing a pipeline that mixes semantic and structural work — distillation, extraction with validation, approval gates, orchestrated multi-step workflows. Catalogs four load-bearing layering patterns (LLM→Code guard, Code filter→LLM, LLM judge + Code enforce, Code orchestrator + LLM worker) with when-to-use, failure modes, and minimal code sketches. Macro-level companion to the micro "which tool for this one task" question.
read-only-instruments
Design pattern for read-only instruments — aggregate readings over an agent's stored state (distributions, compositions, cluster structure) that inform the operator before an intervention. Use when quantifying an open observation before intervening (instrument-first sequencing), when a design floor/threshold would otherwise be guessed, when calibrating an embedding-based reading (three-point scale), or when deciding whether to build OR remove an instrument (signal-first both ways). NOT for per-event audit logs that replay a decision offline (that is replayable-audit-logs) and NOT for metrics that feed gates, ranking, or retrieval — instruments are observability, never intervention.
replayable-audit-logs
Design pattern for observability-by-default in LLM agents — every feature that performs external I/O, calls an LLM, or makes non-deterministic/heuristic decisions ships a replayable append-only JSONL audit log in the same change. Use when adding or reviewing such a feature (the review question "which log answers why, and can we replay it offline?"), when designing a new audit record schema, when a recurring failure needs corpus-driven repair (replay harness, positive/negative ground truth, regression fixtures from real traffic), or when deciding how to store untrusted text in a log.
shadow-mode-validation
Design pattern for shadow-mode validation — running a candidate decision mechanism (typically an LLM judgment) in observe-only parallel with the live path, recording what it WOULD have decided per event, and letting the accumulated record decide enforcement. Use when an unvalidated stochastic mechanism is about to replace or filter a live behavior (a one-way door for output quality), when a selector/classifier/gate has no published reliability evidence for the model class in play, or when designing the isolation, kill-switch, and exit criteria for a shadow deployment. NOT for aggregate readings over stored state (that is read-only-instruments), NOT for the audit-log record schema itself (that is replayable-audit-logs — a shadow log IS one of those logs), and NOT a substitute for unit tests — shadow mode validates decision quality in production traffic, not code correctness.
agent-adoption-triage
Five-question triage for routing AI work to the right architecture — script, algorithmic search, LLM workflow, or autonomous agentic loop — before any agent is built. Use when someone says "let's build an agent for X", when deciding whether a use case actually needs an autonomous agent, when reviewing an agent deployment proposal or an existing agent for accountability risk, or when choosing between a pre-defined LLM workflow and a runtime-directed loop. Covers the Business AI Quadrants, the design/operation Phase axis, gap-bearer naming, the Phase-crossing decision, per-quadrant governance requirements, and an anti-patterns checklist for recovering misrouted work.
llm-as-judge
Design pattern for LLM-as-judge evaluators — binary checks as evidence, one named holistic verdict, no score aggregation. Use when designing or reviewing any LLM-based quality gate, evaluator, judge prompt, or verdict schema; when a judge's rubric scores fluctuate between runs; when you catch yourself asking an LLM for a 1-5 score, averaging check results, or thresholding a satisfaction ratio. NOT for choosing whether a task needs deterministic or semantic processing, and NOT for the architecture-level judge+enforce state-mutation split.
generation-audit
Model-generation-change audit orchestrator — collect the runtime layer (system prompt + tool descriptions) from the live session, cross-check every self-authored asset (rules / CLAUDE.md / skills / agents) against it, classify mismatches as conflict / redundancy / drift, judge each with the intent-evidence-freshness-expiry frame, then hand the evidence to rules-stocktake / skill-stocktake / agent-stocktake for verdicts. Use when a new Claude model generation ships and assets written for the previous one may now conflict with the substrate — 「世代交代したのでハーネスを照合して」「新モデルに合わせて棚卸しして」 "generation audit", "audit my harness against the new model". NOT for — routine single-layer audits (call the stocktakes directly); rule compliance → skill-comply; whole-config GC → config-gc.
agent-stocktake
Audit ~/.claude/agents/*.md (subagent definitions) for description-layer residency cost, body-layer quality, suppression instructions, staleness, and substrate absorption, assigning Keep/Improve/Update/Merge/Demote-to-skill/Dissolve/Retire verdicts. Use when the user says "audit my agents", "agent stocktake", "which agents should I retire or merge", 「agent を棚卸しして」「エージェント定義を見直して」, or when the model generation changed and agent bodies written for the previous one may suppress or over-constrain the current one. NOT for — skill quality → skill-stocktake; rules → rules-stocktake; runtime 層との横断照合 → generation-audit; whole-config GC → config-gc.
git-workflow
この環境で git コマンドを Bash 実行するときの permission 摩擦回避作法。git add / commit / push / status 等を実行する前に参照する。git 同士の && 連結は通るが、cd を git と混ぜると双方許可済みでもプロンプトする（git -C を使う）。commit メッセージに $( ) を含めない。push は sandbox 無効化が必要。
write-prompt
Generate concise LLM prompts using the lightweight prompt-writer agent. Avoids main-model overthinking.
article-stocktake
公開済みZenn/Dev.to記事の実測メトリクスを収集し、内容品質ランクと実測tierの乖離をproject-localに報告する。Use when — 月次または記事2〜3本ごとの受信状況を確認するとき。NOT for — テーマ候補の生成・順位付け、本文改稿、媒体共通の執筆フロー。
publish-article
quality-gate PASS済みの対象稿1本について、Zennのfrontmatter・preview・予約またはDev.toのdry-run・予約、公開索引、push確認を扱うproject-local公開skill。Use when — 対象となるZenn稿またはEN稿自身のglobal quality-gate PASSと著者GOの後。NOT for — 翻訳生成、prose review、security/quality判定、note/Substack投稿。
zenn-format
Zenn記事のfrontmatter、emoji/topics、Zenn固有Markdown記法の正本。Use when — Zenn原稿を作成・検証するとき、本文凍結後にtopics/emoji候補を提示するとき。NOT for — 執筆構成・voice・タイトル判定・公開��否・Dev.to/note/Substack形式。
substack-publishing
完成・レビュー済みのhuman essayをnote/SubstackへHTML pasteで公開し、project corpusへmirrorするlocal workflow。Voice / AI-slop / Title / 出典はglobal writing-ecosystem、翻訳はprose-translationにdefer。Use when — note/Substackへ手動投稿・mirror・配信運用するとき。NOT for — 執筆、翻訳、quality判定。
schedule-publish
記事バッチの公開順序と日程を 4 軸スコアリングで決定し schedule.json に反映する。投稿タイミングの値は zenn-writing.md が正本。
Showing top 100 of 111 skills by quality score. Browse the full catalog at /browse.
Bio shown is the top-scored skill's repo description as a fallback — real GitHub bios land in a future update.