一键导入
retro
Turn session friction, repeated mistakes, and human corrections into minimal rule or skill improvements for approval. 日本語の依頼例:「振り返り」「レトロ」「学びをルール化」「同じミスを防ぎたい」。
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。
菜单
Turn session friction, repeated mistakes, and human corrections into minimal rule or skill improvements for approval. 日本語の依頼例:「振り返り」「レトロ」「学びをルール化」「同じミスを防ぎたい」。
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
Frontend/BFF security for forms and external input, auth, cookies, HTML, URLs, fetch, CORS, webhooks, uploads, secrets, logging, dependencies, and LLM-bound content. 日本語の依頼例:「認証実装」「問い合わせフォーム」「外部API」「webhook」「ログ」「LLMに渡す」「依存追加」。
Repository policy design or changes: CI, releases, dependency or license policy, secrets scanning, protected-branch/PR rules, and performance budgets. Not for routine metadata edits such as a one-line .gitignore change. 日本語の依頼例:「CI/CD設定」「リリース手順」「依存ライセンス」「ブランチ/PR運用」。
Next.js-specific App Router work: Server/Client Components, data fetching, caching, Server Actions, route handlers, metadata, and env vars. Not for framework-neutral styling or tests. 日本語の依頼例:「Next.jsで実装」「Server Action」「RSC」「ルートハンドラ」「キャッシュ」。
Pre-merge verification covering scope, typecheck, lint, affected tests, security, accessibility, and a single pass or fail report. 日本語の依頼例:「出荷前チェック」「マージ前確認」「完成か確認」「PR前チェック」。
React component, hook, state, effect, and rendering conventions. Use for React implementation or re-render and useEffect debugging. 日本語の依頼例:「Reactコンポーネント」「カスタムフック」「状態管理」「再レンダリング」「useEffect修正」。
Accessibility implementation and review for semantic HTML, keyboard use, focus, ARIA, contrast, forms, and screen readers. Use for UI components or accessibility fixes. 日本語の依頼例:「アクセシビリティ対応」「キーボード操作」「フォーカス管理」「スクリーンリーダー対応」。
| name | retro |
| description | Turn session friction, repeated mistakes, and human corrections into minimal rule or skill improvements for approval. 日本語の依頼例:「振り返り」「レトロ」「学びをルール化」「同じミスを防ぎたい」。 |
The rules improve only if what went wrong feeds back into them. This skill is that loop. Output is a PROPOSAL — .agents/** and .codex/** edits need human sign-off (Codex PreToolUse cannot prompt; obtain sign-off before the tool call), so never apply silently.
git log --oneline -20 + PR review comments if accessible — repeated fix-up commits (fix: after feat: on the same files) are rule-gap evidence.description (trigger), not a new rule — say which.frontend-security/a11y and need human security review per governance.For each candidate, output:
tooling)? A machine check beats a prose rule — propose the config change if so.Also propose DELETIONS: a rule nobody has hit in months is a pruning candidate (security/a11y rules are exempt — see governance).
End with the proposals as ready-to-apply diffs and wait for sign-off. After applying any, add a CHANGELOG.md entry.