Claude Code Configuration

My working setup for Claude Code — skills, agents, hooks, slash commands, output styles, and the CLAUDE.md that drives it.

Take what's useful, ignore the rest.

Quickstart

The repo lives at ~/Claude and symlinks into ~/.claude/:

git clone <your-fork> ~/Claude
ln -s ~/Claude/CLAUDE.md      ~/.claude/CLAUDE.md
ln -s ~/Claude/agents         ~/.claude/agents
ln -s ~/Claude/commands       ~/.claude/commands
ln -s ~/Claude/hooks          ~/.claude/hooks
ln -s ~/Claude/skills         ~/.claude/skills
ln -s ~/Claude/output-styles  ~/.claude/output-styles

Then create your own MEMORY.md, settings.json, and the todos/, memories/, projects/ directories. They're gitignored — that's where your real context lives.

Reference

Skills

Coding work:

• browser-tools — Browser automation via Chrome DevTools.
• codebase-documenter — Write READMEs, architecture docs, and code comments.
• commit — Conventional-commit workflow; read before committing.
• frontend-design — Distinctive frontend interfaces; no generic AI aesthetics.
• frontend-philosophy — React/Preact conventions: TanStack Query, no TypeScript, no barrel files.
• improve-codebase-architecture — Consolidate modules, improve testability and AI navigation.
• insecure-defaults — Catch hardcoded secrets, weak auth, permissive config.
• playwright-skill — Browser automation with Playwright.
• prototype — Build a throwaway prototype to flesh out a design.
• quality-code — Type-safe, well-tested, observable TypeScript.
• requesting-code-review — Verify work meets requirements before merging.
• scaffold-astro — Scaffold Astro projects with Bun + Tailwind.
• skill-creator — Create or update skills.
• systematic-debugging — Root-cause-first discipline for hard bugs.
• test-driven-development — Red-green-refactor loop.
• to-issues — Break a plan into tracer-bullet issues.
• to-prd — Turn a conversation into a PRD and publish it.
• triage — Triage incoming issues through a state machine.
• using-git-worktrees — Isolated worktrees for feature work.
• visual-audit — Screenshot and analyze visual composition.

Writing and communication:

• caveman — Ultra-compressed output mode; cuts tokens ~75%.
• compress — Compress memory files (CLAUDE.md, todos) to caveman format.
• writing-clearly-and-concisely — Strunk's writing rules for docs, commits, errors, UI copy.

Integrations:

• bird — Twitter/X via the bird CLI.
• cloudflare — Cloudflare API: DNS, Workers, Tunnels, zones.
• gccli — Google Calendar CLI.
• gdcli — Google Drive CLI.
• hetzner-cloud — Hetzner Cloud infrastructure via hcloud.
• jarvislabs — GPU experiments on JarvisLabs.ai.
• localmaxxing — Local-LLM benchmark and eval leaderboards.

Agents

• dhh-code-reviewer — JavaScript/TypeScript review in DHH's style.
• go-reviewer — Idiomatic Go review against Effective Go.
• react-masters-reviewer — React/Preact review through Linsley, Florence, Dodds, Pocock, Abramov, Browne.
• builder — Executes one engineering task at a time.
• validator — Read-only verification that a task met its acceptance criteria.
• api-frontend-tester — Systematic API testing with curl.
• api-backend-tester — Backend testing with pytest and Python tooling.
• test-coverage — Add tests retroactively to existing untested code.
• docs-fetcher — Pull current API references and gotchas before implementing.

Slash Commands

• /plan — Write an engineering plan to specs/.
• /plan_w_team — Same, with a multi-agent team pass.
• /build — Implement the plan.
• /pickup — Pick up where you left off last session.
• /remember — Capture this session's outcome to memory.
• /question — Answer questions about the project without writing code.
• /generate-visual-plan — Render a feature plan as a visual.

Hooks

Python scripts (run via uv inline deps) wired to Claude Code lifecycle events:

• Session start/end — load context, log sessions.
• Notifications — audio cues when Claude needs input.
• Validators — ruff runs after every Python write.
• Subagent sounds — distinct cues per agent type.
• TTS — ElevenLabs, OpenAI, Piper, pyttsx3, Qwen.

Output Styles

Swap response format on the fly: genui, markdown-focused, table-based, ultra-concise.

What's Gitignored

Path	Why
MEMORY.md	Persistent context, preferences, active state
settings.json	Customize your own
*.local.json	Machine-specific overrides
memories/	Session snapshots
projects/	Living docs for active work
todos/	Task tracking
.env	API keys
logs/, status_lines/	Runtime ephemera

Dependencies

• Claude Code
• uv — hooks run via uv run --script
• PipeWire — pw-play for notification sounds on Linux
• ruff — Python validator, auto-installed via uvx

License

WTFPL