Claude Code Starter Kit

Claude Code Starter Kit

Production-extracted architecture for Claude Code — stop losing context, stop repeating yourself, start shipping.

[](LICENSE) [](https://github.com/martingarramon/claude-code-starter-kit/stargazers) [](https://github.com/martingarramon/claude-code-starter-kit/network/members) [](https://github.com/martingarramon/claude-code-starter-kit/commits/main)

[Quick Start](#quick-start) · [Architecture](#architecture) · [Skills](#skills) · [Examples](#examples) · [Contributing](#contributing)

---

> State of the art — this is a V1 snapshot from March 2026. The patterns below still work as a starting point, and the philosophy ("rules from mistakes, not pre-optimization") still drives my live system. What's evolved since: > - Domain specialization. A single CLAUDE.md is fine for general use, but a real working system specializes — mine focuses on agentic AI and business-automation pipelines, with skills and references partitioned by domain. > - Three-layer pattern by refresh cadence. CLAUDE.md (spec — monthly), SKILL.md (brain — rarely), memory/project_*.md (state — every session). Each fact lives in exactly one layer. The repo's flat 3-layer model is a good entry point; this is the next step. > - Multi-AI orchestration. A coordinator skill that routes work across Claude, Codex, Grok, and Gemini, with per-territory dispatch and a connection registry. > > If you want the current state of the art, see yulicreative.ai.

---

The Problem

Claude Code is powerful, but without structure it:

• Forgets your project rules every session
• Skips verification and ships broken code
• Wastes context window on things it already knows
• Gives inconsistent results across sessions

This starter kit fixes all of that.

What This Is

A 3-layer configuration architecture extracted from months of daily production use with Claude Code. Every rule exists because Claude got it wrong — not because it seemed like a good idea.

┌─────────────────────────────────────┐
│           CLAUDE.md                 │  ← Constitution: runs every turn
│   Rules, verification, planning     │
├─────────────────────────────────────┤
│           Skills/                   │  ← Specialists: activated on demand
│   Debugging, code review, ADRs...  │
├─────────────────────────────────────┤
│           Memory/                   │  ← Persistence: survives sessions
│   Decisions, patterns, commands     │
└─────────────────────────────────────┘

Who Is This For

• Solo developers using Claude Code who want consistent, reliable results
• Teams standardizing how Claude Code behaves across projects
• Power users who've hit the limits of ad-hoc prompting and want structure
• Anyone migrating from Cursor/Copilot who wants to replicate their rules in Claude Code

Not for: building Claude API applications (use the SDK for that). This is for configuring the Claude Code CLI itself.

Quick Start

git clone https://github.com/martingarramon/claude-code-starter-kit.git
cd claude-code-starter-kit
chmod +x setup.sh && ./setup.sh

That's it. The setup script creates the directory structure and copies templates. Non-destructive — won't overwrite existing files.

Architecture

Layer 1: CLAUDE.md — The Constitution

The CLAUDE.md file runs on every Claude Code interaction. It contains:

• Verification rules — "Run before claiming done. State the method first."
• Planning rules — "3+ files = plan mode. No guessing."
• Code quality rules — "No placeholders. Prefer editing existing files."
• Communication rules — "One step at a time. Lead with problems."
• Environment config — Shell, OS, paths, tools
• Context hygiene — When to compact, what to load

Layer 2: Skills — Specialist Workflows

Twelve battle-tested workflows that activate on demand:

| Skill | What it does | When to use | |-------|-------------|-------------| | debugging | Systematic root-cause analysis | Anything breaks or behaves unexpectedly | | code-review | 5-step structured review | Before merging or after major changes | | architecture-decision | ADR process with max 3 options | Technical choices with tradeoffs | | research | Multi-source decomposition | Complex questions needing investigation | | context-management | Context window hygiene | Long sessions approaching limits | | tool-integration | MCP server workflow | Building or connecting external tools | | proposal-writing | Client-first structure | Scoping work for clients | | testing | AI-driven TDD strategy | Starting features, ensuring reliability | | deployment | CI/CD pipeline design | Setting up automated workflows | | migration | AI tool transition guide | Switching from Cursor/Copilot | | performance | Profiling and optimization | Slow code, bottlenecks | | security-audit | Proactive vulnerability scanning | Before deployments, periodic reviews |

Layer 3: Memory — Persistent State

The memory/MEMORY.md file persists across sessions:

• Key decisions with rationale
• Patterns learned from mistakes
• Frequently used commands
• Project status

Supporting Infrastructure

• Hooks — Shell scripts that run at session start, after tool use, on errors
• Examples — Stack-specific CLAUDE.md templates, MCP server reference
• Permissions — .claude/settings.json with safe defaults (deny rm -rf, deny force push)

Examples

Stack Templates

Ready-to-use CLAUDE.md configurations for specific tech stacks:

| Template | Stack | |----------|-------| | python-api.md | FastAPI + SQLAlchemy + PostgreSQL | | saas-webapp.md | Next.js 14 + Prisma + PostgreSQL | | nodejs-typescript-api.md | Express/NestJS + TypeScript + PostgreSQL | | react-frontend.md | React + Vite + TypeScript | | go-microservice.md | Go + Gin + Redis | | data-science.md | Python + Jupyter + Pandas + scikit-learn | | serverless.md | AWS Lambda + SAM + Python/Node.js |

Hook Examples

| Hook | Type | Purpose | |------|------|---------| | session-start.sh | Session Start | Git status, last commit, memory reminder | | post-tool-use.sh | PostToolUse | Error logging, last 50 entries | | pre-tool-use-security.sh | PreToolUse | Block dangerous commands before execution | | notification-slack.sh | Notification | Alert team on deployments or errors | | session-stop-cleanup.sh | Stop | Cleanup temp files, generate session summary | | post-tool-use-autocorrect.sh | PostToolUse | Auto-lint/format after file creation | | subagent-performance.sh | SubagentStop | Log sub-agent duration and outcomes |

MCP Server

A working reference implementation (140 lines, Python) for building your own MCP tools. Includes a 5-minute setup guide.

Design Principles

1. Rules from mistakes, not pre-optimization — Every rule exists because Claude got it wrong
2. Progressive context loading — Load what you need, when you need it
3. Skills over prompts — Structured workflows beat ad-hoc instructions
4. Verification before completion — Always run it, never assume it works
5. Memory is cheap, context is expensive — Persist decisions, load on demand

Comparison

| Feature | This Kit | Bare Claude Code | .cursorrules | Ad-hoc Prompts | |---------|----------|-----------------|--------------|----------------| | Persistent rules | CLAUDE.md constitution | None | Rules file | Copy-paste each time | | Specialist workflows | 12 battle-tested skills | Manual each time | Community rules | Reinvent every session | | Cross-session memory | MEMORY.md system | None | None | None | | Verification protocol | Built-in rules | Hope for the best | Varies | None | | Hook automation | 7 hook examples (all API types) | Manual setup | N/A | N/A | | Stack templates | 7 ready-to-use | Start from scratch | Some available | N/A | | MCP server reference | Working example | Docs only | N/A | N/A |

Roadmap

• [x] ~~More stack templates~~ — 7 templates covering Node.js, React, Go, Python, serverless
• [x] ~~More hook examples~~ — 7 hooks covering all Claude Code API types
• [x] ~~Additional skills~~ — 12 skills including testing, deployment, migration, performance, security
• [ ] Community skill directory (share and discover skills)
• [ ] Self-learning patterns (auto-improve CLAUDE.md from outcomes)
• [ ] MCP ecosystem browser (discover community MCP servers)
• [ ] Team collaboration guide

Contributing

See [CONTRIBUTING.md](CONTRIBUTING.md) for guidelines. We welcome:

• New skills and stack templates
• Hook examples for different workflows
• Bug reports and improvements
• Documentation fixes

License

MIT — see [LICENSE](LICENSE).

Built By

Martin Garramon at Yuli Creative — agentic AI systems and business-automation pipelines.

Need a custom Claude Code setup or a system architect's review of yours? Get in touch.

Source & license

This open-source MCP server is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: martingarramon
• Source: martingarramon/claude-code-starter-kit
• License: MIT

Install and usage instructions live in the source repository linked above.