A high-performance project template for Claude Code, implementing the principles from The Architect's Guide to Agentic Vibe Coding.
A ready-to-clone scaffold that gives you:
- Context Architecture — Layered
CLAUDE.mdfiles that keep Claude oriented without bloating context - Deny-by-Default Permissions — Sensitive operations blocked until explicitly allowed
- Multi-Layered Memory — Enterprise → user → project → local settings hierarchy
- MCP Integration — Pre-configured Model Context Protocol examples for common servers
- Subagent Patterns — Actor and Evaluator agents for code review, security, testing
- Workflow Automation — Explore → Plan → Code → Commit cycle with quality gates
# Clone the scaffold
git clone <your-fork-url> my-project
cd my-project
# Run interactive setup
./scripts/setup.sh
# Start working with Claude Code
claudeThe setup script will:
- Check prerequisites (
claude,jq,git) - Copy
.examplefiles to their real locations - Install git hooks
- Print next steps
.
├── CLAUDE.md # Project memory (edit this for your project)
├── CLAUDE.local.md.example # Template for personal overrides
├── .claude/
│ ├── settings.json # Deny-by-default permissions + hooks
│ ├── agents/ # Subagent definitions
│ ├── skills/ # Reusable skill workflows
│ └── commands/ # Slash commands
├── .mcp.json # Team MCP config (committed)
├── .mcp.local.json.example # Personal MCP config (gitignored)
├── examples/ # Reference configs for all features
├── scripts/ # Setup, hooks, and workflow helpers
└── docs/ # Guides for each system
Claude Code loads configuration in a specific order, with later layers overriding earlier ones:
| Layer | File | Scope | Committed? |
|---|---|---|---|
| Enterprise | /etc/claude-code/CLAUDE.md |
All users on machine | N/A |
| User | ~/.claude/CLAUDE.md |
All your projects | No |
| Project | ./CLAUDE.md |
This project | Yes |
| Local | ./CLAUDE.local.md |
Your personal overrides | No |
See docs/memory-hierarchy.md for details.
The scaffold ships with deny-by-default permissions. Sensitive operations (writing .env, pushing to main, deleting files) are blocked. Safe operations (reading files, running tests, linting) are allowed.
See docs/permission-strategy.md for the full policy.
Four pre-built agents in .claude/agents/:
| Agent | Pattern | Purpose |
|---|---|---|
code-reviewer |
Actor | Reviews code for quality, style, and correctness |
security-reviewer |
Actor | Checks for OWASP top 10 and security issues |
test-writer |
Actor | Generates tests matching project conventions |
evaluator |
Evaluator | Attempts to break code by finding edge cases |
See docs/subagent-patterns.md for usage.
Three reusable skills in .claude/skills/:
- pr-check — Review PRs against a customizable checklist
- gen-test — Generate unit and integration tests
- explore-plan-code — Full Explore → Plan → Code → Commit workflow
Example configurations for common MCP servers in examples/mcp/:
- GitHub (HTTP-based)
- Filesystem (stdio-based)
- Web Fetch
- PostgreSQL Database
- Multi-server combined setup
See docs/mcp-guide.md for setup instructions.
The recommended workflow follows four phases:
- Explore — Understand the codebase before changing it
- Plan — Design the approach, get approval
- Code — Implement with quality gates
- Commit — Conventional commits with verification
See docs/workflow-guide.md for the full guide.
- Edit
CLAUDE.mdto describe your project's architecture, conventions, and rules - Adjust
.claude/settings.jsonpermissions for your tech stack - Configure
.mcp.jsonfor your team's MCP servers - Modify subagents and skills to match your workflow
- Update git hooks for your language/framework
MIT — see LICENSE.