Claude Code Configuration

My working setup for Claude Code. Skills, agents, hooks, slash commands, output styles, and the CLAUDE.md that ties it all together.

This repo lives ~/Claude and it's folders are symlinked into ~/.claude/

Clone it, steal what's useful, make it yours.

What's Here

.
├── CLAUDE.md              # System prompt — personality, philosophy, workflow rules
├── agents/                # Subagent definitions (code reviewers, testers, builders, researchers)
├── commands/              # Slash commands (/remember, /resume, /plan, /diff-review, etc.)
├── hooks/                 # Lifecycle hooks — session start/end, notifications, validators
│   ├── sounds/            # Audio cues for agent start/stop and alerts
│   ├── utils/             # TTS engines, LLM helpers (Anthropic, OpenAI, Ollama)
│   └── validators/        # Post-tool-use validators (ruff linter, file checks)
├── skills/                # 30 skills — from frontend design to SEO audits to PDF manipulation
├── output-styles/         # Response format presets (bullet points, zen master, ultra-concise)
├── scripts/               # Standalone tools (committer, visual audit)
└── docs/                  # Design plans and reference material

What's Not Here

These are gitignored because they're personal or ephemeral:

Path	Why
MEMORY.md	Persistent memory — my context, preferences, active state
settings.json	Claude Code settings — customize your own
*.local.json	Machine-specific overrides
memories/	Session snapshots — what happened, decisions made
projects/	Living docs for active work
todos/	Task tracking
.env	API keys
logs/, status_lines/	Runtime ephemera

Highlights

CLAUDE.md

The core of the setup. Defines personality, technical philosophy (DHH-inspired simplicity), git safety rules, workflow patterns, and coding conventions. Read it — even if you don't use anything else, a well-crafted CLAUDE.md transforms the experience. BUT YOU SHOULD CUSTOMIZE THIS - it's styled for me.

Skills (30)

Each skill is a self-contained capability Claude can invoke. Some I built, some are community skills I've adapted.

Development: frontend-design, frontend-philosophy, scaffold-astro, mcp-builder, web-artifacts-builder, systematic-debugging, playwright-skill

Design: awwwards-design, mk3y-design, ui-ux-pro-max, canvas-design, visual-audit, visual-explainer

Writing & Marketing: copywriting, writing-clearly-and-concisely, doc-coauthoring, brand-identity, marketing-psychology, competitive-analysis

SEO: seo-audit, seo-competitor-analysis, programmatic-seo

Research: mckinsey-research, audit-website

Tooling: pdf, tmux, uv, commit, skill-creator

Agents (18)

Subagents for parallel and specialized work:

• Code reviewers — DHH-style, React/Preact masters, Rust, component analysis
• Team — builder, validator, Next.js expert, React/TypeScript specialist
• Testers — API frontend (curl-based), API backend (pytest), test writer
• Research — system architect, docs fetcher, PRD writer, YouTube API expert, LLM/AI research
• Utilities — meta-agent (creates new agents), training data generator, work completion summary, Hono stack scaffolder

Hooks

Python scripts (using uv inline dependencies) that fire on Claude Code lifecycle events:

• Session start/end — load context, log sessions
• Notifications — audio alerts when Claude needs input
• Validators — ruff linter runs automatically after Python file writes
• Subagent sounds — different audio cues for different agent types starting/stopping
• TTS utilities — ElevenLabs, OpenAI, Piper, pyttsx3, Qwen text-to-speech

Slash Commands (20)

Custom commands beyond the built-ins:

/remember /resume /plan /build /prime /diff-review /fact-check /plan-review /project-recap /web-diagram /question /memories /save-thread /sentient /reconnect and more.

Output Styles

Swap Claude's response format on the fly: bullet-points, ultra-concise, table-based, yaml-structured, markdown-focused, zen-master, tts-summary, genui.

How to Use This

Option 1: Cherry-pick what you want

Copy individual skills, agents, or hooks into your existing ~/.claude/ setup. Each piece is self-contained.

Option 2: Read and adapt

The CLAUDE.md alone is worth studying. The patterns - session memory, slash commands for workflow, lifecycle hooks, work regardless of specific implementation.

After Cloning

1. Create your own MEMORY.md (gitignored — this is your persistent context)
2. Create your own settings.json (gitignored — configure permissions, models, MCP servers)
3. Create todos/, memories/, projects/ directories for your session data
4. Add a .env if any hooks need API keys (TTS, etc.)

Plugins I Use (Not Committed)

Plugins are installed via Claude Code's plugin system (/install-plugin) and live in ~/.claude/plugins/. You'll see references to these throughout the skills list — they're not in this repo, but they're part of the workflow.

Superpowers — The big one. Adds structured workflows for brainstorming, plan writing, TDD, parallel agent dispatch, code review, debugging, and git worktree management. If you only install one plugin, make it this.

Ralph Loop — Autonomous loop that lets Claude keep working without waiting for input at each step. Useful for executing multi-step plans.

Trail of Bits Security — Two plugins:

• insecure-defaults — flags insecure patterns in generated code
• ask-questions-if-underspecified — forces Claude to clarify before guessing

Dependencies

• Claude Code (obviously)
• uv — hooks use uv run --script for zero-config Python execution
• PipeWire — pw-play for notification sounds (Linux; swap for afplay on macOS)
• ruff — Python linting validator (auto-installed via uvx)

Philosophy

This setup follows a few principles:

Convention over configuration. Skills, agents, and commands follow consistent structures. Adding a new one means copying a template and filling it in.

Personal over portable. The gitignored files (memory, settings, todos) are where the real value lives. This repo provides the scaffolding; your context makes it work.

Composable over monolithic. Every piece — a single skill, one hook, a slash command — works independently. Take what serves you, ignore the rest.

License

WTFPL. Do what the fuck you want with it.