claude

Global Claude Code config. Layered scope intent:

Scope	Path	Purpose
User (global)	~/.claude/	Cross-project tools, preferences
Project (shared)	<repo>/.claude/	Codebase-specific, committed
Project (local)	<repo>/.claude/settings.local.json	Per-machine overrides, gitignored

This module ships the global scope.

What's stowed

• CLAUDE.md — global preferences (code style, workflow)
• settings.json — permission allowlist + agent toggles (see below)
• skills/ — hand-picked skills from mattpocock/skills: caveman, grill-me, grill-with-docs, handoff

What's NOT stowed

Runtime state (transcripts, sessions, plugin cache, paste cache, telemetry) stays as real dirs under ~/.claude/. .gitignore uses allowlist pattern so future Claude versions can add new runtime paths without leaking into the repo.

Stow quirk: --no-folding

~/.claude/ is shared with runtime writes, so install.sh passes --no-folding for this module only. Without it, stow would symlink the whole ~/.claude → repo and runtime writes would land in the repo. With it, individual children (settings.json, CLAUDE.md, skills/*) get linked, and ~/.claude stays a real dir.

CLAUDE.md

# Global preferences

## Code style
- No comments unless absolutely necessary. `// TODO:` to flag future work is fine.
- Functional: composable, pure, immutable, single-purpose. Pipelines over imperative loops where idiomatic.

## Workflow
- Before touching unfamiliar code, run `grill-me` or `grill-with-docs` to confirm understanding of the existing model.
- Boy-scout: context-free cleanup only (renames, dead code, obvious extracts). Commit the refactor. Tests green before and after.
- Lay out the full planned commit sequence for the task (big picture) for review before writing code.
- Structural refactors that only make sense in light of the task land as their own commit at the front of the plan (Fowler preparatory refactor).
- Logically atomic commits — one logical unit per commit.
- Hotfix / on-call: skip grill and refactor, fix forward, refactor later.
- Never run `git commit`. Stage as needed, then propose the commit command (HEREDOC for multi-line bodies) and stop. User runs it.

settings.json

{
  "permissions": {
    "allow": [
      "Bash(npm run lint)",
      "Bash(npm run test:*)",
      "Bash(bun run lint)",
      "Bash(bun run test:*)"
    ],
    "deny": [
      "Read(.env)",
      "Read(.env.*)",
      "Read(**/.env)",
      "Read(**/.env.*)"
    ]
  },
  "skipDangerousModePermissionPrompt": true,
  "remoteControlAtStartup": true,
  "agentPushNotifEnabled": true
}

What each entry does

Key	Effect
permissions.allow: Bash(npm run lint), Bash(bun run lint)	Pre-authorize lint runs. Exact-match patterns — no args allowed, intentionally. Lint shouldn't need arguments at this level.
permissions.allow: Bash(npm run test:*), Bash(bun run test:*)	Pre-authorize test runs with any args (:* matches "command + any trailing args"). Lets Claude scope tests (e.g. npm run test -- path/to/file) without prompting.
permissions.deny: Read(.env), Read(.env.*), Read(**/.env), Read(**/.env.*)	Hard block on reading dotenv files anywhere in the tree. Prevents accidental secret exfiltration to a model context. Deny rules override allow rules.
skipDangerousModePermissionPrompt: true	Suppress the one-time confirmation when entering dangerous mode (bypass-all-prompts). Opted in once, globally.
remoteControlAtStartup: true	Enable the remote control feature at session start — lets the claude.ai web app drive the local CLI session.
agentPushNotifEnabled: true	Allow agents to push notifications (desktop / phone via Claude app) when long-running work completes.

Permission pattern syntax cheatsheet

• Bash(cmd) — exact match. No args.
• Bash(cmd:*) — cmd followed by any args. The :* suffix is the documented "command + arbitrary trailing args" pattern.
• Read(path) — exact path match relative to cwd. No ./ prefix.
• Read(glob) — glob match. **/ for recursive, * for single-segment wildcard.

What's not in here

• Hooks. The old Superset notification hooks (gated by $SUPERSET_HOME_DIR) were removed once that tool fell out of use.
• Plugin enablement. The caveman plugin was disabled when the mattpocock caveman skill replaced it. Skill files in skills/ are simpler to manage than plugin install metadata.
• Secrets / per-machine overrides. Belong in ~/.claude/settings.local.json (gitignored). Don't add machine-specific paths or tokens to the stowed settings.json.