Chief ⚔️

English | ไทย

A structured workflow for AI coding agents. Drop it into any project, set your rules once, and stop re-explaining your codebase every chat.

Chief is part of the chief-tribe ecosystem. It uses sage as its behavioral baseline.

Why Chief exists

Every project has context — the decisions from six months ago, the weird workaround, all the “why we do it this way” stuff. It lives in your head. Every new AI chat starts blank, so you re-explain. Then again next chat.

Chief stops that. Give every project the same shape — AGENTS.md for rules, .chief/_rules/ for standards, .chief/milestone-N/ for current work. Agents know where to read and write. Your prompts shrink to one sentence.

→ Why Chief exists

Quickstart

Step 1 — Install skills:

npx skills@latest add thaitype/chief

Select the skills you want. Make sure chief-install is included.

Step 2 — Run /chief-install in your agent:

/chief-install

It asks which coding agent you use, whether to symlink or copy, and whether to include subagents. That’s it.

Step 3 — Bootstrap project context (optional):

/chief-init

Interviews you about your stack and dev commands, writes .chief/project.md. Skip it and write the file by hand later if you prefer.

→ Full tutorial: your first milestone→ Manual install options

Windows users: Symlink mode requires Developer Mode and git config --global core.symlinks true. The install skill auto-detects and falls back to copy mode.

How Chief works

Chief is markdown files in three places:

project/
├── AGENTS.md          ← framework + project rules (highest authority)
└── .chief/
    ├── project.md     ← tech stack, dev commands (written by /chief-init)
    ├── _rules/        ← standards that apply across all milestones
    └── milestone-1/   ← current work: goals, contracts, tasks

.chief/ is created lazily — nothing appears until you need it.

Rules hierarchy: AGENTS.md > .chief/_rules/ > .chief/milestone-N/_goal/. Higher always wins.

→ Rules hierarchy → Directory structure

Working styles

Controlled — review every step

Best for complex projects, unfamiliar domains, team work.

/chief-plan        # grill → goals → contracts → TODO → tasks (approval at each step)
builder-agent: implement task-1 from milestone-1
/chief-retro       # review and capture lessons as rules

Autonomous — let AI drive

Best for prototyping, well-defined goals, solo work.

/chief-autopilot   # reads goals + contracts, runs all tasks
/chief-retro

Mix and match

/chief-plan        # plan with approval gates
/chief-autopilot   # execute the approved plan
/chief-retro

Skills

→ Full skills reference → How to pick the right skill

Agents

→ Subagents reference

Upgrading

# 1. Refresh skills
npx skills@latest add thaitype/chief

# 2. Upgrade framework files
/chief-upgrade

To pin a version: npx skills@latest add thaitype/chief#v4.0.0 / /chief-upgrade v4.0.0

→ How to upgrade

Documentation

Full documentation lives in docs/manual/:

Compatibility

Releases

• v1 — Initial release, Claude Code support. docs
• v2 — Multi-agent support, skills system. docs
• v3 — Rebranded to Chief. chief- skill prefix. Repo moved to thaitype/chief.
• v4 — Skills via npx skills (decoupled from install). Lazy .chief/. New skills: /chief-init, /chief-rule, /chief-grill, /grill-design, /shape-up, /slim-down, /chief-loop, /loop-readiness. answer-verifier-agent replaces deprecated review-plan-agent.

Branches

• release/v1, release/v2 — Stable legacy releases
• main — Latest stable (v4)
• canary — Active development, may be unstable

Development

To test changes locally:

# Install from your branch in a separate test project
npx skills@latest add thaitype/chief# --skill chief-install

# Then test:
/chief-install 

Contributing

1. Fork and branch from canary
2. Make changes
3. Test with the development workflow above
4. PR targeting canary
5. Commit style: type: description (e.g. feat: add kiro support)

Acknowledgements

• /grill-design and /chief-grill originated from mattpocock's grill-me skill
• Multi-agent architecture inspired by vercel-labs/skills