A Claude Code skill that implements a hierarchical multi-agent development workflow. The main agent acts as a project manager — orchestrating task decomposition, sub-agent dispatch, context preservation, Git version control, and human-in-the-loop quality gates across the full software development lifecycle.
Main Agent (Project Manager)
├─ Maintains global state: CLAUDE.md + PROGRESS.md + ADR + Development Log
├─ Decomposes requirements into atomic sub-tasks
├─ Packages minimal context → dispatches sub-agents
├─ Parses structured reports → updates state
├─ Runs mandatory per-node human review gates
└─ Manages Git: branches, commits, merges, tags
│
▼ dispatches
Sub-Agents (pure functions, zero shared context)
├─ Developer — code implementation (worktree isolation)
├─ Tester — test writing, security scanning, performance
├─ Reviewer — code review, quality inspection
├─ Architect — tech selection, architecture design (Plan agent)
└─ Researcher — competitive research, code exploration (Explore agent)
Sub-agent as pure function — Each sub-agent receives a complete, self-contained input package (task + files + constraints + acceptance criteria). No context inheritance. No sibling awareness. Returns a structured report.
Context preservation — Five-file system ensures nothing is lost across sessions:
| File | Role |
|---|---|
CLAUDE.md |
Project brain — architecture, constraints, active tasks |
PROGRESS.md |
Task board — all tasks with status markers |
docs/adr/ |
Decision archive — why each major choice was made |
docs/lessons-learned.md |
Mistake log — rejections, root causes, preventions |
docs/development-log.md |
Shared chronology — timestamped reports from every agent |
Fine-grained commits — One commit per logical unit, never squash an entire sub-task. Each commit tells a single clear story and is independently reviewable and revertable.
Mandatory human gates — Every development node pauses for human review with a generated testing checklist. Approved → merge. Rejected → document lesson → rework → re-review.
| Phase | Sub-Tasks | Human Checkpoint |
|---|---|---|
| Requirements | Research → User stories → Acceptance criteria → Prioritization → Edge cases → Compliance | Stories + priorities confirmed |
| Design | Tech selection → Architecture (C4) → DB schema → API contract → Component tree → Routing | Tech stack + architecture confirmed |
| Development | Bottom-up per module, one feature branch per sub-task | Every node reviewed before merge |
| Testing | Test plan → Unit → Integration → E2E → Edge → Performance → Security (OWASP) → Accessibility | Security scan confirmed |
| Deployment | IaC → Docker → CI/CD → Monitoring → Release runbook | Final pre-release confirmed |
Every development sub-task must pass this gate before merging:
- Review diff — Verify commits are fine-grained, no off-limits files modified
- Generate testing checklist — Happy path, edge cases, integration, error handling, regression risk, code quality
- Present to human — Checklist + commit history + verification steps
- Process response:
- Approved → merge, delete branch, next task
- Rejected → document in lessons-learned, strengthen constraints, re-dispatch on same branch, re-review
- Rejected twice → escalate to user with options
main ← Stable release (merge from develop only)
└─ develop ← Integration branch
└─ feature/<module>-<task-slug> ← One per sub-task
Commit format: <type>(<scope>): <description>
Types: feat | fix | chore | docs | test | refactor
.claude/skills/project-manager/
├── SKILL.md ← Main workflow (460 lines)
└── references/
├── templates.md ← 11 templates for all file types and reports
└── phase-details.md ← Checklists, pitfalls, exit criteria per phase
Invoke in Claude Code by mentioning project startup or management:
- "Start a new project" / "开始新项目"
- "Build a web app for..."
- "Manage my development workflow"
The skill triggers automatically and guides you through the full SDLC with human checkpoints at every critical decision.
MIT