A complete, stack-agnostic setup for Claude Code with a minimal router CLAUDE.md, 5 behavioral rules, 16 custom skills, 3 automation hooks, and the obra/superpowers plugin — all configured globally so every project gets them automatically.

CLAUDE.md is a ~20-line router — not a knowledge dump. It contains compaction recovery instructions and core behavioral pointers. All detailed rules live in ~/.claude/rules/ where they load automatically into context.

Why? Based on How To Be A World-Class Agentic Engineer: skill/agent lists in CLAUDE.md are redundant (they're auto-discovered), and behavioral rules need to survive context compaction.

Rules load automatically from ~/.claude/rules/ into every session. They survive compaction because Claude re-reads them on recovery.

Skills are reusable instruction sets that Claude follows when triggered. They live in ~/.claude/skills/ and are available across all projects.

Superpowers by Jesse Vincent is a complete software development methodology built as composable skills. These auto-activate based on context and enforce disciplined workflows.

Hooks are event-driven scripts that run automatically in response to Claude Code actions. They use Node.js for JSON parsing (no jq dependency).

Runs before every file edit. Blocks writes to:

• .env* files (environment secrets)
• credentials.json, secrets.yaml, etc.
• Lock files (package-lock.json, yarn.lock, Cargo.lock, etc.)
• .git/ internals

Returns exit code 2 with an explanation, which prevents the edit and shows the message to Claude.

Runs after every file edit. Auto-detects the project's formatter:

Walks up from the edited file to find the project root, checks for formatter config files, and runs the first match. Does nothing if no formatter is configured.

Runs before Claude claims completion. Auto-detects the test runner:

If tests fail, returns exit code 1 which blocks Claude from finishing and feeds the failure output back so it must fix the issue first.

Loaded into every Claude Code session. A minimal ~20-line router with compaction recovery instructions, core behavioral pointers, and a reference to ~/.claude/rules/. Skills and agents are auto-discovered — they are not listed here.

Registers:

• 3 enabled plugins (frontend-design, github, superpowers)
• 3 hooks (PreToolUse, PostToolUse, Stop)
• Model preference (opus)
• Custom status line showing project name and git branch

git clone https://github.com/YOUR_USERNAME/claude-code-setup.git
cd claude-code-setup
bash install.sh

1. Copy skills:

   cp -r skills/* ~/.claude/skills/

2. Copy rules:

   mkdir -p ~/.claude/rules
   cp rules/*.md ~/.claude/rules/

3. Copy hooks:

   mkdir -p ~/.claude/hooks
   cp hooks/*.sh ~/.claude/hooks/
   chmod +x ~/.claude/hooks/*.sh

4. Copy global CLAUDE.md:

   cp CLAUDE.md ~/.claude/CLAUDE.md

5. Merge settings (don't overwrite — merge the hooks section into your existing settings.json):

   # Review settings.json and manually merge into ~/.claude/settings.json

6. Install superpowers plugin:

   claude plugin marketplace add obra/superpowers-marketplace
   claude plugin install superpowers@superpowers-marketplace

7. Restart Claude Code for hooks to take effect.

The router architecture keeps idle cost minimal while ensuring rules survive compaction:

With a ~200K token context window (~800K chars), the idle cost is ~0.5% of capacity. The slight increase over the old ~2,100 chars is worth it — rules now survive compaction and enforce consistent behavior throughout long sessions.

Use the built-in meta-skill:

/skill-creator my-new-skill — does X when Y happens

Or manually:

mkdir ~/.claude/skills/my-skill
# Write ~/.claude/skills/my-skill/SKILL.md with YAML frontmatter + instructions
# Skills are auto-discovered — no need to update CLAUDE.md

Remove or comment out the hook entry in ~/.claude/settings.json under the hooks key.

Add disable-model-invocation: true to any skill's YAML frontmatter to make it manual-only.

~/.claude/
├── CLAUDE.md                              # Minimal router (~20 lines)
├── settings.json                          # Hooks, plugins, preferences
├── rules/
│   ├── coding.md                          # Anti-patterns, security, style
│   ├── commits.md                         # Commit frequency, format, push cadence
│   ├── research.md                        # Research/impl separation, anti-sycophancy
│   ├── completion.md                      # Task completion contracts
│   └── project-setup.md                   # Spec-driven dev, session discipline
├── hooks/
│   ├── auto-format.sh                     # PostToolUse formatter
│   ├── block-sensitive-files.sh           # PreToolUse security guard
│   └── test-enforcement.sh               # Stop hook test runner
├── skills/
│   ├── ask-first/SKILL.md                 # Requirement clarification
│   ├── planning/SKILL.md                  # Task decomposition
│   ├── code-review/SKILL.md               # Structured review
│   ├── security-review/SKILL.md           # Security audit
│   ├── tdd/SKILL.md                       # Test-driven development
│   ├── testing/SKILL.md                   # Test writing patterns
│   ├── debug/SKILL.md                     # Root cause analysis
│   ├── figma/SKILL.md                     # Figma design integration
│   ├── refactor/SKILL.md                  # Safe refactoring
│   ├── performance/SKILL.md               # Profiling & optimization
│   ├── api-design/SKILL.md                # API conventions
│   ├── git-workflow/SKILL.md              # Git standards
│   ├── parallel-agents/SKILL.md           # Concurrent subagents
│   ├── document/SKILL.md                  # Documentation generation
│   ├── skill-creator/SKILL.md             # Meta-skill builder
│   └── webmcp/SKILL.md                   # WebMCP agent-friendly tools
└── plugins/
    └── superpowers/                       # obra/superpowers plugin
        └── skills/                        # 14 methodology skills

• obra/superpowers by Jesse Vincent — development methodology plugin
• Trail of Bits — inspiration for the ask-first skill
• Anthropic — official skills format and reference implementations
• Figma MCP Server — official Figma integration for AI coding agents
• WebMCP — Chrome's browser-native API for AI agent tool exposure
• Chrome DevTools MCP — browser automation MCP server