Claude Dev Team

A portable multi-agent development team for Claude Code. Install into any project to get specialized agents for backend, frontend, UX, infrastructure, architecture, QA, and product — with multiple orchestration modes and a one-command onboarding flow.

Quick Start

# 1. Install into your project
./install.sh /path/to/your/project

# 2. Start Claude Code in your project, then onboard:
/dev-setup                              # Reads project, asks questions, prefills context

# 3. Use the team:
/dev-team Build a new feature               # Implement + quick verify
/dev-team-full Ship-quality implementation   # Implement → review → iterate
/dev-grill Stress-test my plan              # Challenge every assumption

Commands

Setup & Planning

Command	What it does
/dev-setup	Project onboarding. Reads docs and code, asks about your scope, tickets, designs, docs platforms. Prefills all agent context files. Run once after installing.
/dev-grill	Stress-test a plan. Grills you one question at a time from every angle. Produces documentation (committed) and implementation plans (local).

Orchestration

Command	What it does
/dev-team	Standard flow. Analyze → delegate → implement → quick verify (lint, types, build). Fast.
/dev-team-full	Thorough flow. Implement → QA review → UX review → Architecture review → design comparison → classify as Ship/Improve/Rethink → iterate fixes (max 3 rounds). Includes visual design comparison against Figma/mockups when available.

Individual Specialists

Command	Role	Writes code?
/dev-backend	Server-side code, APIs, data models	Yes
/dev-frontend	UI components, state, styling	Yes
/dev-ux	Usability, accessibility, copy review	No (recommends)
/dev-infra	Docker, CI/CD, deployment	Yes
/dev-architecture	System design, API contracts, modules	No (recommends)
/dev-qa	Tests, linting, edge cases	Yes
/dev-product	Requirements, acceptance criteria	No (recommends)

When to Use What

Situation	Command
Just installed, first time	/dev-setup
Before a big feature	/dev-grill
Simple single-domain change	/dev-backend, /dev-frontend, etc.
Multi-domain feature (fast)	/dev-team
Ship-quality feature (thorough)	/dev-team-full
Bug fix	/dev-backend or /dev-frontend

Onboarding Flow (/dev-setup)

When you run /dev-setup, it:

1. Reads everything available — CLAUDE.md, README, docs/, package.json, .mcp.json, CI config
2. Asks only what it can't find — your scope (whole project or specific sections?), ticket management (GitHub Issues, Linear, Jira?), design tools (Figma?), doc platforms (Notion?), deployment, decision style
3. Prefills all 7+ context files — so every agent starts with full project awareness
4. Discovers integrations — notes which MCP servers are configured and which might need setup

This means the first /dev-team invocation after setup is fast — agents already know the codebase.

Design Comparison (/dev-team-full)

When UI changes are involved and design references exist, /dev-team-full adds a visual comparison step:

1. Captures a screenshot of the implementation (via Preview MCP, Chrome MCP, or user-provided)
2. Fetches the design reference (Figma MCP, design spec in docs, or provided image)
3. Compares on: layout, typography, colors, states, responsiveness, missing/extra elements
4. Classifies each deviation as Intentional, Bug (route to frontend), or Ambiguous (ask user)

This only runs when designs are available. If none exist, it's skipped with a note.

How It Works

• Skills (.claude/skills/dev-*/SKILL.md) — slash commands that set specialist thinking
• Agents (.claude/agents/*.md) — subagents spawned by orchestrators with restricted tools and worktree isolation
• Context files (.claude/context/<role>.md) — per-role working memory, read at start, written at end, saves tokens on re-exploration

Install & Update

# First install (safe — never overwrites existing files)
./install.sh /path/to/project

# Update templates (replaces agents + skills, preserves context files)
./install.sh --update /path/to/project

Customizing

1. Edit agent prompts — replace generic tech references with your stack
2. Ensure CLAUDE.md is comprehensive — all agents read it first
3. Add reference files — put project-specific patterns in skills/dev-*/references/
4. Add onboarding questions — edit skills/dev-setup/SKILL.md Phase 2, follow the existing pattern

Prerequisites

• Claude Code CLI
• A project with a CLAUDE.md documenting conventions and structure

Infrastructure

No special infrastructure needed. Git worktrees handle parallel agent work on a single machine.

License

MIT