spawning-plan

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

Spawning Plan

Agent生成计划

Design the optimal agent team for the task. Performant, precise, minimal. Docs: https://code.claude.com/docs/en/agent-teams.md

Task:

$ARGUMENTS

为任务设计最优的Agent团队，确保高性能、精准且精简。文档：https://code.claude.com/docs/en/agent-teams.md

任务：

$ARGUMENTS

Step 1: Context Gathering (Silent — no user interaction)

步骤1：上下文收集（静默执行——无需用户交互）

A) Read environment:

```
CLAUDE.md
```
— workflow rules, conventions, constraints

Project manifests —

package.json

pyproject.toml

Cargo.toml

go.mod

, etc.

Directory structure —
```
src/
```
,
```
app/
```
,
```
packages/
```
, test dirs, monorepo indicators

B) Inventory existing agents:

Scan
```
~/.claude/agents/*.md
```
— reuse matching agents instead of creating duplicates

C) Analyze task complexity:

Work type: research, implementation, review, debugging, refactoring
Scope: single-layer vs cross-layer
Parallelism: can work split into independent streams?
Complexity → team size: simple (2 agents), medium (3-4), complex cross-cutting (5-6, max 8)

A) 读取环境信息：

```
CLAUDE.md
```
— 工作流规则、约定、约束条件

项目清单文件 —

package.json

、

pyproject.toml

、

Cargo.toml

、

go.mod

等

目录结构 —
```
src/
```
、
```
app/
```
、
```
packages/
```
、测试目录、单体仓库标识

B) 盘点现有Agent：

扫描
```
~/.claude/agents/*.md
```
— 复用匹配的Agent，避免重复创建

C) 分析任务复杂度：

工作类型：研究、实现、评审、调试、重构
范围：单层 vs 跨层
并行性：工作是否可拆分为独立流？
复杂度→团队规模：简单任务（2个Agent）、中等任务（3-4个）、复杂跨领域任务（5-6个，最多8个）

Step 2: Ask Team Questions (AskUserQuestion Tool)

步骤2：询问团队相关问题（使用AskUserQuestion工具）

Ask 3-5 questions based on Step 1 findings. Not all apply every time — pick what matters.

Team Composition — "For this [work type] on [stack], I'm thinking [N] agents: [role list]. What would you change?" Options: Perfect / Add role / Remove role / Different approach
Coordination — "How should agents work together?" Options: Independent (no messaging) / Team (peer messaging) / Hub-spoke (lead coordinates)
Dependencies — "Work order?" Options: All parallel / Sequential (A→B→C) / Mixed
Models — "Model allocation: opus (research), sonnet (implementation), haiku (scanning). Adjust?" Options: As suggested / All opus / All sonnet / Custom
Agent Reuse (only if matching agents found in Step 1B) — "Found existing
```
[agent-name]
```
that handles [capability]. Reuse it?" Options: Reuse / Create fresh / Both

基于步骤1的结果，提出3-5个问题。无需每次都问所有问题，仅选择相关的即可。

团队构成 — "针对[工作类型]任务，基于[技术栈]，我计划配置[N]个Agent：[角色列表]。您有什么调整建议？" 选项：完美 / 添加角色 / 删除角色 / 更换方案
协作方式 — "Agent之间应如何协作？" 选项：独立协作（无消息交互）/ 团队协作（对等消息交互）/ 中心辐射模式（由主导Agent协调）
依赖关系 — "工作执行顺序？" 选项：全部并行 / 顺序执行（A→B→C）/ 混合模式
模型分配 — "模型分配方案：opus（研究任务）、sonnet（实现任务）、haiku（扫描任务）。是否需要调整？" 选项：按建议执行 / 全部使用opus / 全部使用sonnet / 自定义配置
Agent复用 （仅当步骤1B中找到匹配的Agent时询问） — "发现现有
```
[agent-name]
```
可处理[能力范围]。是否复用该Agent？" 选项：复用 / 创建新Agent / 两者结合

Step 3: Output & Approval

步骤3：输出计划并获取批准

Present clean TEAM PLAN:

undefined

呈现格式清晰的TEAM PLAN：

undefined

TEAM PLAN

Task: [description] Pattern: [independent | team | hub-spoke] Work Order: [parallel | sequential | mixed] Agents: [count]

Teammates

Teammate 1: [Name] ([Role]) Description: [1-2 line expertise and specialization] Model: [opus|sonnet|haiku] Type: [general-purpose | feature-dev:code-X | reuse ~/.claude/agents/X.md] Responsible for: [specific deliverable] Depends on: [— | Teammate N]
Teammate 2: [Name] ([Role]) Description: [1-2 line expertise and specialization] Model: [opus|sonnet|haiku] Type: [general-purpose | feature-dev:code-X] Responsible for: [specific deliverable] Depends on: [— | Teammate N]
...

Teammate 1: [Name] ([Role]) Description: [1-2 line expertise and specialization] Model: [opus|sonnet|haiku] Type: [general-purpose | feature-dev:code-X | reuse ~/.claude/agents/X.md] Responsible for: [specific deliverable] Depends on: [— | Teammate N]
Teammate 2: [Name] ([Role]) Description: [1-2 line expertise and specialization] Model: [opus|sonnet|haiku] Type: [general-purpose | feature-dev:code-X] Responsible for: [specific deliverable] Depends on: [— | Teammate N]
...

Research (injected into agent prompts)

[key finding or best practice 1]
[key finding or best practice 2]

undefined

[key finding or best practice 1]
[key finding or best practice 2]

undefined

Final Approval (AskUserQuestion Tool)

最终批准（使用AskUserQuestion工具）

"Launch this team?"
- Deploy & Save — spawn agents and save as reusable skill
- Deploy Once — spawn agents, one-time
- Adjust — change something (iterate plan)
- Cancel — abort

Deploy & Save → save team as skill at

~/.claude/skills/<team-name>/SKILL.md

for future use via

/<team-name> [task]

. Saved skill skips planning, bakes in agent definitions, uses

$ARGUMENTS

for task input.

Adjust → ask what to change → regenerate plan → ask again. Loop until approved.

Deploy Once → spawn immediately, no save.

Cancel → stop.

"是否启动该团队？"
- Deploy & Save — 生成Agent并保存为可复用Skill
- Deploy Once — 生成Agent，仅单次使用
- Adjust — 修改计划（迭代调整）
- Cancel — 终止操作

Deploy & Save → 将团队保存为Skill，存储路径为

~/.claude/skills/<team-name>/SKILL.md

，后续可通过

/<team-name> [task]

调用。已保存的Skill会跳过规划步骤，内置Agent定义，使用

$ARGUMENTS

作为任务输入。

Adjust → 询问需要修改的内容→重新生成计划→再次获取批准。循环直至获得批准。

Deploy Once → 立即生成Agent，不保存。

Cancel → 停止操作。

Spawning Execution

生成执行逻辑

Based on chosen pattern:

Independent: parallel Task tool calls, one per agent
Team: TeamCreate → TaskCreate per agent → Task tool with
```
team_name
```
→ TaskUpdate for dependencies
Hub-spoke: TeamCreate with lead agent (opus) that delegates via SendMessage

For detailed agent prompt structure, see references/agent-prompts.md.

根据所选模式执行：

独立模式：并行调用Task工具，每个Agent对应一次调用
团队模式：先执行TeamCreate→为每个Agent执行TaskCreate→调用带
```
team_name
```
参数的Task工具→针对依赖项执行TaskUpdate
中心辐射模式：创建包含主导Agent（opus模型）的团队，由主导Agent通过SendMessage指令分配任务

关于Agent提示词的详细结构，请参考references/agent-prompts.md。