# Agent Team

> Work as an agent team. Send non-trivial work to sub-agents; keep the main thread for planning, acceptance, smoke-testing, and communication. A QA inspector is mandatory and holds veto power — a FAIL must be fixed and re-reviewed. Triggers when about to do heavy multi-file edits in the main thread, take on a task spanning many files/modules, or say "I'll just do this myself". The prime goal is to keep the main context lean and avoid overflow. Agent Team 协作:非平凡任务派子 agent,主线程只做规划/验收/冒烟测试/沟通。监理(QA)必选且有否决权。触发:将做 3+ 文件编辑/跨多模块任务,或想"我自己干完就好"时。首要目标是精简主上下文。

- Skill: `yli769227-jpg/agent-team` (Agent Skill, multi-file: 2 files)
- Install (CLI): `npx skillmds@latest add yli769227-jpg/agent-team`
- Raw SKILL.md: https://api.skillmd.com/api/skills/yli769227-jpg/agent-team/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: AI & ML
- Author: yli769227-jpg (https://skillmd.com/u/yli769227-jpg)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/yli769227-jpg/agent-team

---


# Work As An Agent Team / Agent Team 协作

> Prime directive: keep the main context lean. A bloated lead thread is how work derails.
> 首要原则:精简主上下文。主线程一臃肿,工作就脱轨。

## The discipline / 纪律

1. Send non-trivial work to a sub-agent team instead of unrolling huge edits in the main conversation.
   非平凡任务派子 agent,而不是在主对话里展开大量编辑。
2. The main thread does only: plan, accept, smoke-test, communicate.
   主线程只做:规划、验收、冒烟测试、沟通。

## The core roles / 核心角色

- **Engineer** — writes code, makes the thing run. / **工程师**:写代码,让项目跑起来。
- **QA Inspector** — finds the cracks; verifies happy path, error path, and boundaries. / **监理**:找裂纹,验证正确/错误/边界路径。
- **Lead (main thread)** — plans, coordinates, accepts. / **主线程**:规划、协调、验收。
- Add specialist agents (Research, Security, …) as the task needs. / 可按需追加 Research、Security 等专职 agent。

## QA principle / 监理原则

- QA is **not optional** — every non-trivial task gets reviewed before "done".
  监理**不是可选项**——任何非平凡任务完成后必须验收。
- QA's verdict has **veto power**. A FAIL must be fixed and re-reviewed.
  监理结论有**否决权**。FAIL 必须修复后重新验收。

## When this triggers / 触发时机

- About to start a main-thread task that will edit 3+ files / span multiple modules.
- About to take on a task that reads 5+ files / generates large amounts of code.
- About to declare a non-trivial task done (it must pass QA first).

## Sub-task must-include checklist / 派子任务必含清单

When the lead hands work to a sub-agent, the task description MUST spell out all four — this is the single source of truth the other disciplines point back to:
主线程派子任务时,任务描述必须写明这四项——这是其余纪律在"派子任务"场景下的统一出口:

1. **Verification step** — how to confirm it works (test-is-truth). / **验证步骤**——怎么确认跑通(test-is-truth)。
2. **Key-path logging** — connect / state change / error / request-response (log-first). / **关键路径日志**——连接/状态变化/错误/请求响应(log-first)。
3. **Build after each file edit** (incremental-build). / **每编一个文件就 build**(incremental-build)。
4. **Re-read the latest file before editing** — don't trust a cached copy (no-dead-code). / **编辑前重读最新文件**——不信缓存副本(no-dead-code)。

## Done criterion (verifiable) / 完成判据（可验证）

✅ Non-trivial work ran through sub-agents, and a QA inspector reviewed it (happy + error + boundary paths) and returned PASS before "done" was claimed.
⚠️ QA returned PASS-with-caveats — surface the caveats.
❌ The lead thread did everything itself and/or skipped QA. Route it through the team and review.

## Worked examples / 实战反例

Real before/after cases for this discipline live in [EXAMPLES.md](./EXAMPLES.md) — read them before you act.
本纪律的真实 before/after 反例见 [EXAMPLES.md](./EXAMPLES.md) —— 动手前先对照。