菜单
Audit alignment between AGENTS.md and repo-local shared skills under .agents/skills. Invoke explicitly only; do not auto-trigger on generic agent or skill maintenance requests.
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。
| name | agent-maintainer |
| description | Audit alignment between AGENTS.md and repo-local shared skills under .agents/skills. Invoke explicitly only; do not auto-trigger on generic agent or skill maintenance requests. |
このリポジトリの agent 運用ルールと shared skill レイアウトの整合を点検し、必要な整理を行うための skill である。
対象は repo ローカルの以下に限定する。
AGENTS.md.agents/skills/.claude/skills/.codex/skills/.agents/plan/repo 外の ~/.codex/...、MCP server 設定、editor 設定は対象外とする。
AGENTS.md または CLAUDE.md を読み、repo 共通ルールを確認する.agents/skills/, .claude/skills/, .codex/skills/ の構造を確認するAGENTS.md と各 skill の間で、重複・矛盾・古い運用記述を洗い出すAGENTS.mdAGENTS.md には参照だけを残すAGENTS.md を lean に保つ
.agents/skills/<skill>/ に揃える.claude/skills/<skill> と .codex/skills/<skill> を正本ディレクトリ symlink に揃えるAGENTS.md の責務分離を保つ.agents/plan/ の命名規則や参照導線が planning skill と整合していることを確認するSkill Directory Workflow のような shared rule は AGENTS.md を正本とするplanning skill の責務であり、この skill はその定義の整合だけを見るPrevent dev-environment values (origins, ports, hosts, absolute paths, secrets) from leaking into build artifacts, generated code, or production output. Use whenever writing or reviewing build scripts / codegen, adding a URL / port / hostname / origin literal, branching on dev vs prod, configuring CORS or crossorigin attributes, or when generated files (*.gen.*, dist output, feeds, sitemaps) embed any absolute URL. Trigger proactively — this class of bug ships silently and is only noticed in production.
Working discipline for hard tasks — decompose, verify, and decide what to do next. Use whenever a task is non-trivial: the approach is not obvious, it spans multiple steps or files, it involves debugging an unknown cause, an irreversible or production-facing change, or the user says it is hard/complex/stuck. Also use when starting work that a strong model (Opus 4.8+) is delegated precisely because it is difficult. Do not wait for the user to ask for "planning" — apply this protocol proactively at the start of any such task.
Create and iterate on implementation plans under .agents/plan/ so multiple agents (Claude / Codex / 他) can review each other's drafts. Use when entering plan mode, drafting an implementation plan, or reviewing an existing plan.
Create install scripts under ./install/install-<name>.sh following this dotfiles repo's build/install conventions (brew vs self-build, pkg/local layout, OS branching).
Sync and harden command permissions between .claude/settings.json and .codex/rules/default.rules in this dotfiles repo.