macrohard-teams
Claude Code plugin that gives you a 4-agent build team. Tell it what to build — PM writes the PRD, engineer implements, QA tests.
Install
From marketplace (recommended)
# 1. Add the marketplace
/plugin marketplace add tumblecat44/macrohard-teams
# 2. Install the plugin
/plugin install macrohard-teams@macrohard-teams
# 3. Activate (run once after install)
/reload-plugins
Use
/macrohard-teams:build 유저 인증 시스템 만들어줘
The leader agent:
- Decides if a team is needed (small tasks skip the team)
- Has PM write a PRD with acceptance criteria
- Has engineer implement based on the PRD
- Has QA write and run tests against the acceptance criteria
- If tests fail, sends fixes back to engineer (max 3 retries)
- Reports what was built
Check progress:
/macrohard-teams:status
Team
| Agent | Model | Writes to | Enforced by |
|---|
| leader | Opus | orchestration only | never implements |
| product-pm | Sonnet | docs/product/, docs/specs/ | PreToolUse hook |
| eng-lead | Opus | src/, package* | PreToolUse hook |
| qa-tester | Sonnet | tests/, e2e/ | PreToolUse hook |
Boundaries are physically enforced — agents that try to write outside their directories get blocked with exit code 2.
How it works
Wave-based execution
Tasks run in dependency-based waves, not fixed stages:
- Independent tasks (zero file overlap) → parallel (proven 1.75x speedup)
- Dependent tasks → sequential (PRD before implementation, implementation before tests)
Hooks
| Hook | Event | Purpose |
|---|
| enforce-boundary.sh | PreToolUse (Write/Edit) | Blocks agents from writing outside their directories |
| enforce-commit.sh | PreToolUse (Bash) | Requires Conventional Commits format |
| log-decision.sh | PostToolUse (Write/Edit) | Logs every file change to decisions.jsonl |
| teammate-gate.sh | TeammateIdle | Blocks agents from going idle without deliverables |
| task-complete-gate.sh | TaskCompleted | Logs task completions |
| inject-context.sh | SubagentStart | Reminds agents of their boundaries on startup |
Repository structure
macrohard-teams/
├── .claude-plugin/
│ ├── marketplace.json # Marketplace catalog (source: ".")
│ └── plugin.json # Plugin manifest (v2.0.0)
├── agents/ # 4 agent specs
│ ├── leader.md
│ ├── product-pm.md
│ ├── eng-lead.md
│ └── qa-tester.md
├── hooks/
│ └── hooks.json # 6 hook events
├── scripts/ # 6 enforcement scripts
├── skills/
│ ├── build/SKILL.md # /macrohard-teams:build
│ └── status/SKILL.md # /macrohard-teams:status
└── README.md
Research basis
Design decisions are backed by 18 research sources across 8 iterations with zero contradictions:
- Anthropic Official Docs — 3-5 teammates optimal, file locking prevents conflicts
- Google Research — sequential tasks degrade -70% with more agents
- Nicholas Carlini (Anthropic) — tests are the primary control mechanism
- Addy Osmani — planner-worker-judge hierarchy
- Alex Lavaee — Cursor + Anthropic converged on same 5 primitives independently
License
MIT
tumblecat44
