AgentSkillsCN

claude-dev-team

采用 Agent Teams 实现多智能体开发工作流。支持四种模式：计划模式（架构师与项目经理共同探讨 → plan.md）、开发模式（开发者、测试人员与评审人员迭代开发 → 代码）、全周期模式（计划 → 审批关口 → 开发）以及自动化模式（计划 → 开发，无需审批关口）。当任务受益于协作式智能体角色与同行间的消息传递时，可使用此模式。

SKILL.md
--- frontmatter
name: claude-dev-team
description: "Multi-agent development workflow using Agent Teams. Supports four modes: plan (architect + PM debate → plan.md), dev (developer + tester + reviewer iterate → code), full (plan → approval gate → dev), and auto (plan → dev, no gate). Use when tasks benefit from collaborative agent roles with peer messaging."
license: MIT
compatibility: "Requires Claude Code with CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1. Context7 MCP server is bundled via plugin .mcp.json and starts automatically."
allowed-tools: Read Grep Glob Bash Task TaskCreate TaskUpdate TaskList TaskGet Write Edit AskUserQuestion TeamCreate SendMessage TeamDelete WebSearch WebFetch
metadata:
  author: claude-dev-team
  version: "1.0.0"

Claude Dev Team

Multi-agent development workflow with four modes. Pick one based on the user's needs:

ModeWhen to use
planNeed architecture/design before coding
devHave an approved plan, ready to implement
fullEnd-to-end with user approval gate between plan and dev
autoEnd-to-end without approval gate

Before executing any mode, read references/WORKFLOW.md for detailed step-by-step instructions, spawn prompts, and output templates.

Architecture

code
Plan Phase (plan/full/auto)     Dev Phase (dev/full/auto)
  Lead (You)                      Lead (You)
  ├── architect  [teammate]       ├── developer  [teammate]
  ├── prod-mgr   [teammate]      ├── tester     [teammate]
  └── researcher [subagent]       ├── reviewer   [teammate]
                                  └── researcher [subagent]
         │                                │
         └──── plan.md (handoff) ─────────┘

Teammates message each other directly (Architect↔PM, Dev↔Tester↔Reviewer). Researcher is a subagent — Lead relays results.

Roles

Researcher (subagent — spawn via Task without team_name)

Research specialist for doc lookups. Queries Context7 for library docs, searches web for best practices, returns structured findings with code examples. Bundled as agents/researcher.md in this plugin — Context7 MCP is auto-configured via .mcp.json.

Architect (teammate — plan phase)

Designs architecture: components, interfaces, file changes, data flow, testing strategy. Debates tradeoffs with PM. Messages design to lead and PM.

Product Manager (teammate — plan phase)

Validates architecture against requirements. Challenges design with concerns. Produces verdict: APPROVED or NEEDS_REVISION with specifics.

Developer (teammate — dev phase)

Implements tasks from plan. No stubs, no TODOs. Matches existing patterns. Iterates with tester on failures, reviewer on issues.

Tester (teammate — dev phase)

Writes and runs tests matching existing patterns. Messages developer with specific failures + root cause. Max 3 fix cycles, then escalates to lead.

Reviewer (teammate — dev phase)

Reviews changed files for completeness, correctness, security, quality, plan adherence. Validates review with /council (quick quality for routine, review security or review architecture for critical concerns). Scans for stubs. Messages developer with file:line + fix suggestions. Max 3 cycles.

Rules

  • •One team at a time — cleanup plan-team before starting dev-team
  • •Teammates debate directly (Architect↔PM, Dev↔Tester, Dev↔Reviewer)
  • •Researcher is always a subagent — Lead relays results
  • •Plan.md is the single source of truth and handoff artifact
  • •Every task declares depends_on; parallel within waves, sequential between
  • •Verify build between waves
  • •Avoid file conflicts between parallel tasks
  • •Testing + review are mandatory quality gates
  • •Always cleanup team before finishing
  • •Plan mode: do NOT implement — only plan
  • •If stuck — ask user, don't loop