claude-template

A production-ready .claude/ folder template that turns Claude Code into a coordinated engineering team — 3 agents, 12 workflow skills, and strict quality gates.

Inspired by addyosmani/agent-skills and garrytan/gstack.

---

What this is

Most people use Claude Code as a single generic assistant. This template gives it structure.

Three specialist agents. Twelve workflow skills. Always-on rules. Shared context. A process that scales from solo to team.

Every skill encodes senior engineering discipline with anti-rationalization tables and verification gates. The manager orchestrates, the engineer builds, and the reviewer ensures nothing ships broken.

---

The team

Agent	Role
manager	Orchestrates the department. Routes work to the right skill or agent. Runs daily executive summaries. Thinks in budget, capacity, and risk. Reports to you.
engineer	Full-stack implementation. Architecture, APIs, database, frontend, CI/CD, infrastructure. TDD with 100% coverage. Domain-driven design. Performance-measured.
reviewer	Code review, security audit, E2E sign-off, process improvement. Threat modeling (STRIDE). Error logging. Retrospectives. Evolves agent files when patterns repeat.

---

How it works

Every initiative follows this sequence:

Idea
 → /project:spec (define what + why)
 → /project:design (UI specs, if applicable)
 → /project:plan (architecture + task breakdown)
 → /project:build (engineer implements with TDD)
 → /project:review (code + security review)
 → /project:e2e (E2E sign-off)
 → /project:ship (land branch)
 → /project:retro (log + improve)

No skipping steps. The manager enforces the sequence.

---

Slash commands

Command	What it does
/project:spec	Define what to build — use cases, RICE prioritization, acceptance criteria
/project:plan	Architecture design + task breakdown
/project:design	UI/component design specs, mobile review
/project:build	Incremental implementation with TDD
/project:test	Test strategy and coverage enforcement
/project:review	Pre-merge code review — critical issues first, then informational
/project:security	Security audit — threat model, OWASP Top 10, CVE check
/project:e2e	E2E quality assurance and release sign-off
/project:ship	Land a reviewed, signed-off branch
/project:retro	Sprint retrospective + agent improvement pass
/project:daily-report	Manager daily executive summary
/project:investigate	Root cause debugging — no fixes without diagnosis

---

Always-on rules

Applied across every agent, every session:

• tech-stack — languages, frameworks, databases, infrastructure (fill in before first session)
• code-style — naming, formatting, TypeScript strictness, import order
• testing — 100% coverage (statements, branches, functions, lines), TDD, no I/O in unit tests
• api-conventions — REST naming, response envelope, status codes, pagination, versioning
• security — input validation, secrets management, PII handling, auth/authz gates
• ux-guidelines — spacing scale, typography, mobile rules, interaction patterns

---

Non-negotiables

• 100% unit test coverage on all new and modified code — CI blocks merge if not met
• Code review before every merge
• E2E sign-off before every release
• Security review for anything touching auth, data, or external APIs
• Every error logged before it's fixed
• Recurring errors (2+) trigger an agent file update — the reviewer owns this

---

Key patterns

Borrowed from the best in the space:

From agent-skills:

• Anti-rationalization tables — every skill preemptively counters excuses for skipping steps
• Gated verification — require evidence (passing tests, clean builds), not "it seems right"
• Six core operating behaviors baked into the build skill

From gstack:

• "Boil the Lake" — AI makes completeness cheap, so always do the complete thing
• Atomic commits per fix in QA workflows
• Workflow-shaped skills over role-shaped agents

---

Setup

Before your first session, fill in:

File	What to add
.claude/rules/tech-stack.md	Your languages, frameworks, databases, infrastructure
CLAUDE.md	Build commands for your project

That's it. The AI derives status from git log and keeps decisions in memory/progress.md.

---

Directory structure

your-project/
├── CLAUDE.md              — project identity + build commands
└── .claude/
    ├── agents/            — 3 specialist agents (manager, engineer, reviewer)
    ├── skills/            — 12 workflow skills (spec, plan, build, review, ship, etc.)
    ├── rules/             — 6 constraint files (tech-stack, code-style, testing, etc.)
    ├── memory/            — lean project memory (progress.md)
    ├── docs/adr/          — Architecture Decision Records
    └── settings.json      — tool permissions

No sprint trackers, no action item tables, no retrospective directories. The AI reads git history for status and updates a single memory file for decisions and context that git can't capture.

---

Install

Copy the .claude/ folder and CLAUDE.md into your project root:

git clone https://github.com/ianchan0817/claude-template.git
cp -r claude-template/.claude your-project/
cp claude-template/CLAUDE.md your-project/

Then fill in tech-stack.md and your build commands. You're ready.

---

Settings

.claude/settings.json controls what Claude can and cannot do in your project.

Denied by default: rm -rf *, force push, git reset --hard, DROP/truncate on databases, reading .env files, decrypting secrets.

Adjust the allow list to match your actual toolchain (bun, pnpm, cargo, go, docker, etc. are already included).

---

License

MIT