/dev — AI-Assisted Multi-Agent Development SOP for Claude Code
A custom skill for Claude Code that turns Claude into a Tech Lead coordinating multiple AI Worker Agents through a complete software development workflow — from requirements alignment to PR merge.
Why /dev?
Most AI coding tools give you a smart autocomplete. /dev gives you an engineering process.
| Without /dev | With /dev |
|---|---|
| Claude writes code directly in chat | Claude acts as Tech Lead — never writes code itself |
| No structure, easy to lose track | 6-phase SOP from PRD to merge |
| Single-threaded, one thing at a time | Multiple Worker Agents develop in parallel worktrees |
| You catch conflicts at merge time | Pre-coding conflict scan catches overlaps before they start |
| Security review is manual | bandit + pip-audit / npm audit run mandatorily before every Review |
How It Works
You: /dev I want to build a Todo app with user auth and task management
Claude (Tech Lead)
│
├─ Phase 0 ── Classify request → New Project
├─ Phase 1 ── Align on PRD (2-round confirmation)
├─ Phase 2 ── Architecture decisions + GitHub Issues
│
├─ Phase 3 ── Spawn Worker Agents (parallel worktrees)
│ ├─ Worker A: Auth module
│ ├─ Worker B: Task CRUD API
│ └─ Worker C: Frontend components
│
├─ Phase 3.5 ── QA Agent static verification
├─ Phase 4 ── Code Review + merge (7-item checklist)
└─ Phase 5 ── Retro + next iteration loop
What It Does
/dev is a multi-agent software development SOP covering:
- Phase 0 — Request classification & routing (New Project / New Feature / Bug Fix / Emergency Hotfix / Architectural Change / Refactoring)
- Phase 1 — Product alignment (uses your existing PRD, or generates one in two rounds)
- Phase 2 — Architecture decisions + task decomposition + GitHub Issue creation
- Phase 3 — Multiple Worker Agents developing in parallel (each in an isolated worktree)
- Phase 3.5 — QA Agent static verification
- Phase 4 — Code Review + merge (7-item structured checklist with mandatory veto conditions)
- Phase 5 — Retro + next iteration loop
Key Features
- Tech Lead iron rule: main conversation never writes code directly — all changes go through Worker Agent → PR → Review
- 6-category counterexample self-check: Worker Agents must cover Null / Empty / Boundary / External failure / Concurrency / Malicious input before submitting
- Pre-coding conflict check: Worker Agents scan other open Issues for file overlap before starting
- Post-merge PR coordination: after every merge, scans open PRs and notifies branches that need rebase
- Security gate:
bandit+pip-audit/npm auditrun mandatorily before Review - Database migration guard: direct DDL operations (ALTER TABLE / DROP COLUMN) trigger mandatory veto
Requirements
| Tool | Notes |
|---|---|
| Claude Code | Anthropic official CLI, login required |
GitHub CLI (gh) | For Issues, PRs, repos — run gh auth login first |
| Git | Version control |
Installation
Option 1: Script (recommended)
# macOS / Linux — English version
bash install.sh --lang en
# macOS / Linux — Chinese version
bash install.sh --lang zh
# Windows PowerShell — English
.\install.ps1 -Lang en
# Windows PowerShell — Chinese
.\install.ps1 -Lang zh
Option 2: Manual
Copy all files from en/commands/ (or zh/commands/) into ~/.claude/commands/, keeping the directory structure:
~/.claude/commands/
dev.md
dev/
phase1.md
phase2.md
phase3.md
phase4.md
worker-new.md
worker-fix.md
qa-agent.md
PROJECT_CONTEXT_TEMPLATE.md
Usage
In Claude Code, type:
/dev [optional description]
Examples:
/dev I want to build a Todo app with user registration and task management
/dev
Claude will automatically classify the request type and enter the corresponding flow.
Repository Structure
claude-dev-skill/
├── README.md # This file (English)
├── README.zh.md # Chinese version
├── LICENSE
├── install.sh # macOS/Linux installer
├── install.ps1 # Windows installer
├── en/ # English skill files
│ └── commands/
│ ├── dev.md
│ └── dev/
│ ├── phase1.md ~ phase5.md
│ ├── worker-new.md
│ ├── worker-fix.md
│ ├── qa-agent.md
│ └── PROJECT_CONTEXT_TEMPLATE.md
└── zh/ # Chinese skill files
└── commands/
└── dev/ ...
Scope
Best suited for:
- New small-to-medium web backend / API projects
- Feature modules with an existing PRD that need systematic development
- Parallel multi-feature development where you want to avoid merge conflicts
Not suited for:
- Projects requiring production deployment (no DevOps/deployment capability)
- Complex database migrations with large amounts of existing data
- High compliance security requirements (finance, healthcare)
Contributing
Issues and PRs are welcome. If you have ideas for new phases, agent roles, or language support, open an Issue first to discuss.
License
MIT — see LICENSE