Claude Code skills, rules, and per-language project bundles.

Layout

Two install modes

Global (machine-wide)

Symlinks skills and generic rules into ~/.claude/. Run once per machine.

make install       # symlink skills and rules into ~/.claude/
make uninstall     # remove the symlinks
make list          # show what's installed

Skills and generic rules apply to every Claude Code session on this machine.

Per-project language bundles

Copies a language-specific ruleset into a target project. Re-run to refresh.

make install-elisp PROJECT=~/code/my-elisp-thing
# or, explicit:
make install-lang LANG=elisp PROJECT=~/code/my-elisp-thing

make list-languages   # show available bundles

What gets installed:

• .claude/rules/*.md — project-scoped rules (language-specific + verification)
• .claude/hooks/ — PostToolUse validation scripts
• .claude/settings.json — permission allowlist + hook wiring
• githooks/ — git hooks (activated via core.hooksPath)
• CLAUDE.md — seeded on first install only (use FORCE=1 to overwrite)
• .gitignore — appends personal-override entries (deduped)

The install is re-runnable. Running it again refreshes files in place; personal tweaks live in .claude/settings.local.json and are not touched.

MCP servers (user scope)

Registers MCP servers globally (user scope) so every Claude Code project sees them. Reads structure from mcp/servers.json (placeholders ${VAR}), decrypts secrets from mcp/secrets.env.gpg via gpg-agent, expands the placeholders, then registers anything not already present in claude mcp list. Idempotent — re-running is safe.

make install-mcp     # decrypt + register everything in servers.json
make uninstall-mcp   # remove every server listed in servers.json
make check-mcp       # dry-run drift report (no decryption, no writes)

check-mcp classifies each server as ok (in both), MISSING (configured but not registered — run install-mcp), or EXTRA (registered but not configured — usually intentional manual additions like the claude.ai web servers). Exit code is non-zero only on MISSING, since EXTRA entries are often deliberate.

What lands on disk during install-mcp:

• mcp/gcp-oauth.keys.json (mode 600) — extracted for google-calendar-mcp
• ~/.config/google-docs-mcp/{personal,work}/token.json (mode 600) — per-profile OAuth tokens for @a-bonus/google-docs-mcp

Secrets never touch disk in plain form outside the OAuth artifacts above. The .gpg file is the source of truth; rotate via gpg --edit-key and re-encrypt. See mcp/README.org for the full pipeline.

Available languages

Add more by creating languages/<name>/ with the same structure.

Bundle structure

Each language bundle under languages/<lang>/ follows:

languages/<lang>/
├── CLAUDE.md              # project instructions template (seed only)
├── claude/                # copied into <project>/.claude/
│   ├── rules/*.md
│   ├── hooks/*.sh
│   └── settings.json
├── githooks/              # copied into <project>/githooks/
│   └── pre-commit
└── gitignore-add.txt      # lines appended to <project>/.gitignore

Design principles

• Authoritative source: .claude/ and githooks/ overwrite on every install. If you edit them in-project, your changes will be lost on next install. Put per-project customizations in .claude/settings.local.json (gitignored) or project-specific files outside .claude/.
• CLAUDE.md is precious: it's the one file with project-specific prose, so install never overwrites it unless FORCE=1.
• Portable paths: hooks use $CLAUDE_PROJECT_DIR (Claude Code sets it) with a script-relative fallback. No hardcoded usernames or paths.
• Idempotent: re-running install is always safe. No state beyond file contents.