メニュー
Audit and clean AI coding tool settings — permissions, plugins, hooks, cross-layer conflicts. USE WHEN review permissions, clean settings, audit config, plugin conflicts, hook audit, settings cruft.
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
SOC 職業分類に基づく
Create and validate skills for forge modules. USE WHEN create skill, new skill, write skill, validate skill, check skill, skill structure, skill conventions.
Create, validate, evaluate, and iterate skills for forge modules. USE WHEN create skill, new skill, write skill, validate skill, check skill, skill structure, skill conventions, test a skill, run skill evals, benchmark a skill, skill not triggering, optimize skill description. Not for adopting community skills (AdoptArtifact) or shipping artifacts downstream (PublishArtifact).
Create, validate, or place behavioral rules in forge modules. USE WHEN create rule, new rule, write rule, capture rule, add a rule, graduate a correction or convention to a rule, validate rule, rule conventions, where does this rule belong, or the user runs /Rule.
Scaffold a workshop project: a VCS spine (git + jj colocate + entire + forge hooks), the private/public/assets flow layout with a .tlp manifest, a .claude/ that mounts the knowledge vault, and on-demand satellites (vault folder-note, data dir, private remote). USE WHEN starting a new project, initializing a folder for work, adding a vault note or data dir or remote to a project, archiving a project, or the user runs `project` / `atlas`.
Collaborative ideation before implementation — explore context, clarify intent, propose approaches. USE WHEN starting creative work, new features, design decisions, or modifying behavior.
Write a formal design spec from brainstorming output or requirements. USE WHEN design approved, write spec, formalize design, spec document.
| name | SettingsMaintenance |
| version | 0.1.0 |
| description | Audit and clean AI coding tool settings — permissions, plugins, hooks, cross-layer conflicts. USE WHEN review permissions, clean settings, audit config, plugin conflicts, hook audit, settings cruft. |
Audit and clean AI coding tool settings files. Covers permission entries, plugin configuration, hook wiring, and cross-layer conflicts.
AI coding tools layer settings across multiple files. Claude Code uses 4 files with inheritance:
| Layer | File | Committed | Scope |
|---|---|---|---|
| Global shared | ~/.claude/settings.json | No (personal) | All projects |
| Global local | ~/.claude/settings.local.json | No (personal) | All projects, overrides global shared |
| Project shared | .claude/settings.json | Yes (git) | This project, shared with collaborators |
| Project local | .claude/settings.local.json | No (gitignored) | This project, personal overrides |
Precedence (highest wins): project local > project shared > global local > global shared.
At the global level, both files are personal — the settings.json / settings.local.json split is redundant. The local file is designed for project-level use where settings.json is committed and settings.local.json holds personal overrides.
| Subskill | Trigger | Content |
|---|---|---|
| ClaudePermissions | "permissions", "allowed commands", "settings cruft" | @ClaudePermissions.md |
| ClaudePlugins | "plugins", "plugin conflicts", "disable plugins" | @ClaudePlugins.md |
| ClaudeHooks | "hooks", "hook config", "dispatch" | @ClaudeHooks.md |
| ClaudeBaseline | "baseline", "global settings", "user scope", "bootstrap" | @ClaudeBaseline.md |
| GeminiSettings | "gemini settings", "audit gemini" | @GeminiSettings.md |
| Full audit | "full audit", "review settings", "clean everything" | All of the above |
All subskills follow this 6-phase workflow:
jq . <file> > /dev/null for JSON validity. Count before/after entries. Spot-check that critical tools remain permitted.Run these regardless of which subskill is active:
| Check | What to look for |
|---|---|
| Env var duplication | Same key+value in both global and project settings.json — project inherits from global |
| Permission scattering | Same entry in both settings.json and settings.local.json at the same level |
| Plugin enable/disable conflict | Global disables a plugin but project re-enables it (or vice versa) |
| Redundant local file | At global level, settings.local.json entries that already exist in settings.json |
permissions.defaultMode: "auto" routes every tool call through an AI classifier that blocks anything irreversible, destructive, or aimed outside the trusted environment. Configure that classifier with the autoMode block, whose entries are prose natural-language rules, not tool patterns:
autoMode.environment: trusted infrastructure (repos, buckets, domains, local services)autoMode.soft_deny: destructive actions, blocked unless the user's explicit, specific intent clears themautoMode.hard_deny: unconditional boundaries, intent and allow exceptions do not applyautoMode.allow: exceptions to soft blocksPrecedence: hard_deny > soft_deny > allow > explicit intent. Include the literal "$defaults" in any list to keep built-in rules. The classifier also reads CLAUDE.md. Validate with claude auto-mode config and claude auto-mode critique.
This is the soft, AI-judged layer; the hard layer is permissions.deny (before the classifier, unoverridable) and command-guard hooks. permissions.allow/deny/ask use structured Tool(specifier) syntax, not prose.