本文へ移動
cccskills
無料GitHub で公開

run-skill-create

実行して新規Skillを端から端まで作りたいとき、複数Gateを通した品質保証付きフローを起動したいときに使う。

インストール方法を見る

含まれるファイル(17)

  • SKILL.md24.0 KB
  • prompts/R1-elicit.md10.2 KB
  • prompts/R2-gate-review.md5.9 KB
  • prompts/R3-governance-decide.md6.1 KB
  • references/gate-templates.md2.8 KB
  • references/handoff-schema.json713 B
  • references/resource-map.yaml2.6 KB
  • references/skill-brief-schema.json709 B
  • schemas/build-trace.schema.json2.5 KB
  • schemas/findings.schema.json2.3 KB
  • schemas/handoff.schema.json1.9 KB
  • schemas/rubric-merge.schema.json2.2 KB
  • schemas/skill-brief.schema.json11.8 KB
  • scripts/evaluate-create-gates.py2.5 KB
  • scripts/resolve-brief-to-category.py7.9 KB
  • scripts/validate-intake-publish-ready.py4.7 KB
  • workflow-manifest.json14.6 KB

SKILL.md(原文)

インストールする前に、エージェントに与えられる指示の中身を確認できます。

Pre-choice usable artifact execution

Purpose & Output Contractの最小の実成果物をmain contextで作成する。effect別のparse/open・secret・irreversible・corrupt guardだけを実行し、現物path・digest・開き方を提示してからaccept-as-is/light/standard/detailedを記録する。accept-as-isはその場でhandoff完了とし、後続sectionを実行しない。

Post-choice selected improvement execution

以下の既存workflow・goal-seek・評価・修正sectionはlight/standard/detailedが記録されてsemantic_evaluator_startedへ遷移した場合だけ実行する。release/exhaustiveは別の明示eventを必要とする。

run-skill-create

Runtime root contract

runtime_root_policy: host-skill-path の製品別root解決、cwd推測禁止、prompt継承は ref-cross-platform-runtime の共有正本 をそのまま適用する。

Phase 2 移行後は plugins/harness-creator/skills/ が正本、.claude/skills/ は symlink/deploy target。plugin同梱schema/prompt/scriptは上記Runtime root contractで解決したabsolute PLUGIN_ROOT から参照し、repo-root cwdや plugins/harness-creator/... からplugin rootを推測しない。生成対象repoを操作するroot-level gateだけは、明示したproject root/cwdを別入力として使う。

Purpose & Output Contract

ユーザー要望 → skill-brief.json → Skill生成 → plugin/marketplace 登録判定 → P0 lint → 設計評価 → パラダイム評価 → governance 承認 をゲートあり自動連鎖で実行する端から端まで orchestrator。各 Step/Gate の機械可読定義は workflow-manifest.json、責務別プロンプトは prompts/*.md、データ契約は schemas/*.schema.json を参照。

入力: topic (任意), mode ∈ {create, update}, brief_path (任意・E2 直接消費), handoff (任意・brief_path 併給時の parity preflight) 出力:

  • plugins/harness-creator/skills/<skill_name>/ 一式 (SKILL.md + references/ + scripts/)
  • eval-log/skill-build-trace.json (schemas/build-trace.schema.json 準拠)
  • 共通基盤の場合は plugins/harness-creator/.claude-plugin/plugin.json 登録差分
  • eval-log/docs/<NN>-<timestamp>.json (評価結果、schemas/findings.schema.json 準拠)。deprecated: 27章 §3.1 の eval-log/<plugin>/<skill>/<gate>/<run-id>/ (repo root 基準) へ移行
  • eval-log/handoff-<step>.json (schemas/handoff.schema.json 準拠)
  • 完了レポート (日本語本文、パラメーター名のみ英語)

完了条件: P0 lint pass + evaluator JSON pass (--fast 低リスク ref/wrap は evaluator: N/A 理由必須) + (solo_operator_mode 下) LLM-reviewer pass + 各遷移の明示的gate decision。

禁則: ゲート判定なしに次フェーズへ進まない。人間承認gateはユーザー承認を必須とし、solo自動承認はmanifestの全条件を機械評価して得た明示的gate decisionだけを許す。P0 lint 失敗の自動修正は禁止 (根本原因をユーザー提示)。evaluator/governance reviewer は同一 context 評価禁止 (必ず context:fork)。詳細は ## Key Rules。

起動モード

  • 引数なし: Step 1 (run-skill-elicit) が起動、対話で topic を確定。フィールド意味は schemas/skill-brief.schema.json (詳細は references/skill-brief-schema.json)。
  • --brief-path 指定あり (E2 直接消費): 上流 run-plugin-dev-plan の handoff-run-plugin-dev-plan.json を render-skill-brief.py で決定論射影した skill-brief.json を Step 1 の対話ヒアリング (run-skill-elicit) を skip して直接 Step 1 成果物に採用する (再ヒアリングなし)。詳細契約は prompts/R1-elicit.md CONST_014。--handoff 併給時は build dispatch 前に python3 "$PLUGIN_ROOT/scripts/check-route-component-parity.py" <handoff> を実行し exit0 (routes↔inventory 一致) を確認する (CONST_015、非0 で停止)。PLUGIN_ROOT は共有Runtime root contractで解決済みのabsolute rootとする。
  • Notion 指定あり: topic / 引数に --page-url または --page-id が含まれる場合、Step 1 は skill-intake の publish 完了証跡を必須入力とする。output/<hint>/notion-log.json.status=="published"、notion-publish-result.json.page_id、notion-url.txt が揃い、指定 page と一致するまで Step 2 build へ進まない。
  • --fast: 1ファイル変更/<=30行/kind ∈ {ref,wrap}/evaluator pair 不要を全て満たす場合のみ軽量フロー (Step 4b/5 skip)。判定は機械決定:
    python3 "$PLUGIN_ROOT/skills/run-skill-create/scripts/evaluate-create-gates.py" \
      --skill-name "$SKILL_NAME" --kind "$KIND" --brief eval-log/skill-brief.json --fast
    
    条件不一致時は黙って通常フローに戻す。

Key Rules

  1. ゲート前で必ず判定する: 人間承認gateは AskUserQuestion (prompts/R2-gate-review.md) でユーザー承認を得る。solo自動承認はmanifestの全条件が機械的にtrueになった場合だけ明示的PASS decisionとして記録し、条件未充足なら停止する。

  2. 子スキルへの委譲: 各フェーズは独立 Skill を Skill tool で起動 (workflow-manifest.json の delegateSkill)。本スキルは制御のみ。

  3. 失敗時の停止: P0 lint fail または evaluator FAIL なら停止し findings 提示。

  4. context:fork: evaluator/governance reviewer は必ず context:fork で起動 (Sycophancy 防止)。

  5. handoff 保存: 各ゲート通過時に eval-log/handoff-<step>.json を schemas/handoff.schema.json 準拠で残す。PostCompact hook で復元。

  6. plugin/marketplace 登録は確認後: legacy manifest.json は build-manifest-registration-plan.py の提案 → Gate 2.5 承認 → --apply の順。新形式 plugin.json plugin のルート marketplace.json / bundles.json 登録は scripts/validate-plugin-completeness.py --fix(append-only・冪等・書込後自己再検証)が担い、人間が PR diff で最終承認する(破壊的変更を含まない append-only ゆえ生成フロー内で自動実行可)。 Codex 面は run-codex-plugin-package に委譲する。Claude plugin envelopeを持つ全pluginを対象に、新規/既存を状態から判定する同一upsertが .codex-plugin/plugin.json と .agents/plugins/marketplace.json を同期し、手書きの二重管理はしない。Codex固有値は .codex-plugin-overrides.json へ明示する。 これに対し CL-4c の build-plugin-release.py は承認境界の外側 (=承認後) に置く。理由は validate-plugin-completeness.py --fix と性質が違うため: write_version() は plugin.json の version 行を原文 in-place 置換し、そこから公開 marketplace.json (sync_public_marketplace_version()) と .codex-plugin/plugin.json (sync_codex_manifest_version()) を追従させる 既存行の書き換えであり (さらに _run() が bumped 非空のとき regenerate_config_version_lock() で config-version-lock.json も再生成する)、append-only の安全性論拠がそのまま使えない。よって生成フロー内で自動実行してよいのは読み取りのみの --check に限る。version を実際に上げる実行は Gate 2.5 承認後に、--only <plugin> を必ず付けて行う — _run() は only = set(args.only) or None なので無指定は全 plugin 走査となり、本フローと無関係な変更済み plugin まで同じ実行で patch bump されるため、--only は範囲を宣言する必須引数として扱う。

  7. resource-map 先読み: references/resource-map.yaml を最初に読み、必要ファイルのみ open。

  8. 日本語成果物ゲート: brief の output_language=ja と parameter_language_exception=true を既定とし、本文・レビュー・完了レポートを日本語に保つ (パラメーター名・JSON キー・CLI 引数は英語)。

  9. prompt 形式: 新規 prompt は Markdown (.md) 既定。prompts/<R-id>-<slug>.md で plugins/prompt-creator/skills/run-prompt-creator-7layer/references/seven-layer-markdown-template.md を写経して生成。YAML は legacy のみ許容 (新規禁止、P0 lint で warn)。

End-to-End Flow

[Step 1 elicit] run-skill-elicit ─→ skill-brief.json ─[Gate 1]─▶
[Step 2 build]  run-build-skill  ─→ skill-build-trace.json
[Step 3 manifest-register] [Gate 2.5] [Step 3.5 bundle-register]
[Step 3.55 codex-platform-sync] run-codex-plugin-package (Codex manifest + repo marketplace upsert)
[Step 3.56 codex-platform-check] 同じ入力をcheck-onlyで再検査
[Step 3.6 local-marketplace-register] build-local-marketplace.py (非配布 plugin の install 経路)
[Step 3.7 plugin-release-record] build-plugin-release.py --only <plugin> (内容 hash 記録 / patch 採番。Gate 2.5 承認後・--only 必須。Key Rule 6)
[Step 4a p0-lint] (fail→Step 2、最大3周) ─[Gate 2 diff]─▶
[Step 4a.5 pkg-check] run-plugin-package-check (条件: kind==plugin-composition or 新規 plugin 横展開)
[Step 4b design-evaluate] (context:fork) ─→ findings
[Step 5 elegant-review] (条件: new or >30 行, context:fork) ─[Gate 3]─▶
[Step 6 governance] (solo_operator_mode 自動承認) ─[Gate 4]─▶
[Step 7 report]

依存・entryHook/exitHook・resourceIds・fatal_exit_codes は workflow-manifest.json 参照。

ゴールシーク実行

ゴール (Goal)

ユーザー要望から生成された <skill_name>/ 一式が、全 P0 lint pass・evaluator JSON pass・(solo_operator_mode 下) LLM-reviewer pass を満たし、4 ゲート全承認済みで、登録判定・handoff・完了レポートまで完結した状態になっている。

目的・背景 (Why)

新規/更新 Skill を「端から端まで」品質保証付きで送り出すための制御層。各フェーズは独立 Skill へ委譲し、本スキルはゲート制御とハンドオフ整合のみを担う。固定手順では入力 (topic/mode/--fast)・lint 失敗・FAIL 差し戻しなど実行時文脈に脆いため、未達ゲートを都度埋める。

完了チェックリスト (Checklist)

  • eval-log/skill-brief.json が schemas/skill-brief.schema.json 準拠で生成され、Gate 1 承認済み <!-- CL-1 -->
  • Notion 指定ありの場合、python3 "$PLUGIN_ROOT/skills/run-skill-create/scripts/validate-intake-publish-ready.py" --dir output/<hint> --page-url <url> が exit 0。未公開・page_id 不一致・URL 欠落なら Gate 1 で停止し、skill 本体生成へ進んでいない <!-- CL-2 -->
  • <skill_name>/ 一式 (SKILL.md + references/ + scripts/) が Skill(run-build-skill, args=[skill_name, kind, --mode={mode}]) で生成され、eval-log/skill-build-trace.json が schemas/build-trace.schema.json 準拠・章 coverage 全 PASS/N/A/skip 理由付き <!-- CL-3 -->
  • 新規 plugin の場合 python3 ${HARNESS_ROOT:-.}/scripts/validate-plugin-completeness.py --fix 実行済みで、marketplace.json plugins[] + bundles.json (bundle_targets) へ append-only 登録され、検出モード (validate-plugin-completeness.py) が exit 0。プロジェクト固有 (横展開しない) は未登録理由がレポートに記録されている <!-- CL-4 exempt: 登録の運用操作項目。validate-plugin-completeness.py が機械検査し評価 criteria の対象外 -->
  • Claude plugin envelope を持つ場合 同一upsertを実行し、sync-plugin-platforms.py --all --check が exit 0。.codex-plugin/plugin.json と .agents/plugins/marketplace.json の両方が差分に含まれる <!-- CL-4d exempt: Codex 配布 projection を generator と回帰テストが機械検査 -->
  • 新規 plugin の場合 python3 ${HARNESS_ROOT:-.}/scripts/build-local-marketplace.py 実行済みで、marketplaces/local/.claude-plugin/marketplace.json が再生成され --check が exit 0。公開 marketplace は distributable:false を載せられない (MK-004) ため、非配布 plugin を手元の Claude Code へ install する経路はこれだけ。CL-4c を実行するなら本項は不要 (build-plugin-release.py は pending が 1 件でもあれば local marketplace を再生成する)。単独で要るのは pending が空で local marketplace 側だけ drift しているときに限る <!-- CL-4b exempt: 登録の運用操作項目。build-local-marketplace.py --check が機械検査し評価 criteria の対象外 -->
  • python3 ${HARNESS_ROOT:-.}/scripts/build-plugin-release.py --only <plugin> 実行済みで --check が exit 0 (書き込みを伴う実行は Gate 2.5 承認後・--only 必須。Key Rule 6)。marketplace install は version 単位の copy なので、version を据え置いたまま内容を直すと install 済み plugin へ届かない。内容 hash 対応表への記録 (新 plugin) と patch 採番 (既存 plugin の変更) がここで済む <!-- CL-4c exempt: 登録の運用操作項目。build-plugin-release.py --check が機械検査し評価 criteria の対象外 -->
  • 他 plugin リソースを呼ぶ場合 .claude-plugin/bundles.json の現行 bundle (skills-full / skills-intake) のうち plugin.json.bundle_targets で宣言した対象へ登録済み (上記 --fix が bundle_targets を読み自動 append)。不要なら理由がレポートにある (理由なき未登録は rubric 違反) <!-- CL-5 exempt: 登録の運用操作項目。validate-plugin-completeness.py が機械検査し評価 criteria の対象外 -->
  • (legacy manifest.json 形式のみ) plugin/marketplace 登録が Gate 2.5 承認後 --apply 済み (build-manifest-registration-plan.py) <!-- CL-6 exempt: legacy 経路の条件付き運用項目 -->
  • workflow-manifest.json の phases[id=p0-lint].commands 全件 (lint-manifest-contents 含む) が exit 0。TODO/未展開 {{...}}/英語仮文の残存なし (パラメーター名除く) <!-- CL-7 -->
  • create/update 共通で audit-capability-parity.py --plugin plugins/<plugin> --json が PASS。Claude 固有 command/agent/hook は Codex skill 代替または理由・replacement 付き omission contract を持つ。SKILL/promptのplugin-root付き実行行はowner SKILLに runtime_root_policy: host-skill-path とRuntime root本文契約を必須とし、bare CLAUDE_PLUGIN_ROOT、cwd推測、literal placeholderを残さない。env unset + foreign cwdでもabsolute owner SKILL.md pathからmanifest祖先を解決できる <!-- CL-14 -->
  • Gate 2 で git diff <skill_path> + build-trace を提示し承認済み <!-- CL-8 -->
  • assign-skill-design-evaluator (context:fork) の eval-log/docs/<NN>-<timestamp>.json (schemas/findings.schema.json 準拠) が FAIL 残存なし <!-- CL-9 -->
  • 新規 or >30 行変更時、run-elegant-review (context:fork) で C1-C4 全 PASS。PASS 時 eval-log/pattern-feedback.json に pattern_ref_candidates/new_patterns/mass_production_risk を提案保存 <!-- CL-10 -->
  • governance 承認済み: 4 条件 (solo=true/安定版凍結/newly_failing=0/LLM-reviewer pass) を workflow-manifest.json の phases[id=governance].auto_approve_conditions で機械評価し、全充足で自動承認。条件値の正本は repo-root references/governance-params.json (27章§11 パラメータ正本、gitignore。未配備なら references/governance-params.json.example から provision。不在時は graceful degrade=手動承認フロー)。不充足なら run-skill-rubric-governance 経由 (prompts/R3-governance-decide.md R3)。Gate 4 承認済み <!-- CL-11 -->
  • 各ゲート通過時に eval-log/handoff-<step>.json が schemas/handoff.schema.json 準拠で保存されている <!-- CL-12 -->
  • 完了レポート (下記項目: mode/gates_passed/creator_kit_registration/evaluator_result/elegant_review/governance/open_questions) が日本語本文で提示されている (パラメーター名のみ英語) <!-- CL-13 exempt: レポート提示の運用項目。完了レポート契約は本文で規定済み -->

ゴールシークループ

正本 ../run-build-skill/references/goal-seek-paradigm.md の 6 ステップ (現状評価/手順生成/実行/検証/Anchor Step/反復) に従う。本スキル固有の差分:

  • 未達評価の単位はゲート: workflow-manifest.json の gate_order 順に「未承認」とみなして都度埋める (Gate 番号順の直書き禁止、順序正本は manifest)。人間承認gateは AskUserQuestion (prompts/R2-gate-review.md) で承認を取り、solo governanceだけは全 auto_approve_conditions の機械評価PASSを明示decisionとして記録する。
  • 委譲先 (子 Skill): run-skill-elicit / run-build-skill / assign-skill-design-evaluator / run-elegant-review / run-skill-rubric-governance。Notion 指定ありの非技術者 intake は skill-intake 完了証跡を先に検証する。本スキルは制御のみ、各子が自設計書を参照。
  • context:fork 必須: evaluator / elegant-review / governance reviewer は必ず fork で起動 (Sycophancy 防止)。
  • 差し戻し: P0 lint fail または evaluator/elegant FAIL なら run-build-skill 再実行へ戻す (最大 3 周)。--fast 判定・elegant 起動判定は scripts/evaluate-create-gates.py で機械決定 (条件不一致は黙って通常フロー)。
  • lint 自動修正禁止: 根本原因をユーザー提示し LLM 判断で勝手に直さない。
  • 各フェーズ phase=id・entryHook/exitHook・dependsOn・fatal_exit_codes は workflow-manifest.json 参照。プロンプト R1/R2/R3 は prompts/*.md。

Gotchas

  1. Gate skip 禁止: 「次へ」を自動推測しない。人間承認gateは明示確認、solo governanceは全自動承認条件の機械評価PASSが必須。
  2. 同一 context 評価禁止: evaluator/governance reviewer は必ず context:fork (Sycophancy 防止)。
  3. lint 失敗時の自動修正禁止: 根本原因をユーザー提示。LLM 判断で勝手に直さない。
  4. mode=update 時の改名: run-skill-rename に委譲。本スキル対象外。
  5. context 予算: 31 章全部を読まない。本スキルは 05/06/07/13/23/25 章のみ参照。
  6. handoff 保存: 各ゲート通過時に schemas/handoff.schema.json 準拠で必ず保存。
  7. manifest 二重管理禁止: 手書き追加後も lint-manifest-contents.py を必ず通す。

品質ゲート: Elegant Review Protocol

新規/更新/プロンプト改善時は Elegant Review Protocol を適用する。本プロトコルの正本は plugins/harness-creator/skills/run-elegant-review/SKILL.md (Phase 1→2→3 で C1-C4 全 PASS を確認、大規模設計は必須、軽微修正は --fast と整合可)。結果は Step 5 findings に紐付け eval-log/ に残す。

Additional Resources

references/resource-map.yaml を最初に読む (machine-readable と人間向け資料を一覧化)。主要参照:

  • workflow-manifest.json — Step/Gate/Phase の機械可読定義 (entryHook/exitHook/dependsOn/delegateSkill)
  • schemas/skill-brief.schema.json — Step 1→2 渡し正本スキーマ
  • schemas/handoff.schema.json — Gate 通過時 handoff 共通形式
  • schemas/findings.schema.json — orchestrator handoff envelope (evaluator 集約結果の封筒)。Step 5 producer 出力の正本は ../run-elegant-review/schemas/findings.schema.json
  • schemas/build-trace.schema.json — Step 2 emit する章別 coverage 形式
  • schemas/rubric-merge.schema.json — L0/L1/L2 rubric deep-merge 物質化形式
  • prompts/R1-elicit.md / prompts/R2-gate-review.md / prompts/R3-governance-decide.md — R1/R2/R3 責務別プロンプト
  • references/gate-templates.md — Gate 確認質問テンプレ (人間向け詳細手順)
  • 子スキル: run-skill-elicit, run-build-skill, run-plugin-package-check, assign-skill-design-evaluator, run-elegant-review, run-skill-rubric-governance, run-skill-rename
  • 設計書: 05/06/07/11/13/23/25 章

レビュー

まだレビューはありません。使ってみた感想をお寄せください。

同じリポジトリのスキル

概要と使いどころ

app-excellence

無料日本語概要

アプリ・Webサービス・業務ツールの制作を、証拠分析→確定的な初稿→要件定義→体験設計→機能分解→実装→品質ゲート→リリース判定の手順ゲート制で進めるスキル。質問で要件を埋めるのではなく、依頼文・既存コード・資料・ログから最有力案を選び、動く成果物を先に出して差分で改善する。ユーザーが「アプリを作りたい」「ツールを作って」「〜を自動化したい」「システム化したい」「要件定義」「仕様を決めたい」「UI/UXを設計して」「体験を良くしたい」「機能を洗い出して」「MVPを決めたい」「品質チェック」「リリースしていいか判断して」などと言ったら必ず使用する。外部データの取込・マスタ・締め処理を伴う業務システムでは references/data-lifecycle.md も併用する。新規・既存を問わずアプリの制作・改善・リニューアルで使用する。

daishiman/harness-dev102026年10月10日 更新

app-orchestrator

無料日本語概要

Webアプリの準備→要件定義→設計→実装→公開→品質ゲートを実行する内部オーケストレーター。Claude Codeの /build-app /improve-app、Codexの $build-app / $improve-app から明示的に委譲された場合、custom agent起動時、または利用者が $app-orchestrator を明示した場合だけ使用する。一般のアプリ相談から暗黙起動しない。

daishiman/harness-dev102026年10月10日 更新

run-extract-blueprint が生成した章別ブループリントの忠実性を独立 context で評価したいとき、事実/推測区別と粒度と被覆を検証し draft_hash に束縛した PASS/FAIL verdict をローカル品質ゲート (C01 の周回内の受入判定・差し戻し) へ渡したいときに使う。

daishiman/harness-dev102026年10月10日 更新

assign-briefing-evaluator

無料日本語概要

確認点で利用者が見る深さを選んだあと打ち合わせ資料を作り手とは別の目で確かめたいとき、正確さ・言葉・見た目・シンプルさの指摘を資料の編集なしで受け取りたいときに使う。

daishiman/harness-dev102026年10月10日 更新

生成した handout 資料が初心者に伝わるか読みやすさレビューを依頼したいとき、独立 context のレビュアーから指摘と根拠つきの verdict を回収したいときに使う。

daishiman/harness-dev102026年10月10日 更新

Notion ページを描画する直前に粒度を検証したいとき、info-collector-agent ページと同等の section 充足度を section_canonical_map 基準で機械検証したいときに使う。

daishiman/harness-dev102026年10月10日 更新

daishiman のスキルをすべて見る

このスキルの問題を報告する