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

Skill	What it does
/chief-init	Bootstrap .chief/project.md via interview
/chief-plan	Plan a milestone: grill → goals → contracts → tasks
/chief-autopilot	Run a milestone autonomously
/chief-grill	Deep stateful stress-test; spawns answer-verifier-agent per question
/chief-rule	Capture a decision as a permanent rule in _rules/
/chief-retro	Retrospective + lesson learned +_rules/ update
/grill-design	Stateless design stress-test with self-critique
/shape-up	Turn a fuzzy idea into a scoped spec (top-down)
/slim-down	Cut a scope that's too large into a phase-sized piece
/dump-commit	Quick clean commit with auto-generated message

→ Full skills reference → How to pick the right skill

Agents

Agent	Role
chief-agent	Plans, orchestrates, delegates — does not write code
builder-agent	Implements tasks, runs unit tests, commits
tester-agent	Integration/E2E validation — only when you request it
answer-verifier-agent	Background verifier spawned by /chief-grill

→ 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/:

Section	Content
Tutorial	Your first milestone, end to end
How-to guides	Install, upgrade, pick a skill, write rules
Reference	Skills, agents, directory structure, rules hierarchy
Explanation	Why Chief exists, pre-coding first, three-agent model

Compatibility

Coding agent	Integration
Claude Code	CLAUDE.md → AGENTS.md symlink + .claude/ symlinks
GitHub Copilot	.github/agents/ symlinks or copies
Cursor, Windsurf, Kiro, Codex, Aider, Amp, Gemini CLI	Reads AGENTS.md natively (untested ⚠️)

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. 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#<your-branch> --skill chief-install

# Then test:
/chief-install <your-branch>

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