返回技能列表

create-team

pjt222
更新于 2 days ago
2 次查看
17
2
17
在 GitHub 上查看
automationdesign

关于

The `create-team` skill generates a new multi-agent team composition file, adhering to the agent-almanac template and registry conventions. It handles purpose definition, member selection, coordination patterns, and automated README creation for defining coordinated workflows. Use it when composing agents for complex, recurring collaborative tasks or review processes.

快速安装

Claude Code

推荐
主要方式
npx skills add pjt222/agent-almanac -a claude-code
插件命令备选方式
/plugin add https://github.com/pjt222/agent-almanac
Git 克隆备选方式
git clone https://github.com/pjt222/agent-almanac.git ~/.claude/skills/create-team

在 Claude Code 中复制并粘贴此命令以安装该技能

技能文档

Create a New Team

Define multi-agent team composition. Coordinate two or more agents for tasks needing multiple perspectives, specialties, or phases. Resulting team file integrates with teams registry. Can be activated in Claude Code by name.

When Use

  • Task needs many perspectives one agent cannot give (e.g., code review plus security audit plus architecture review)
  • Need recurring collaborative workflow with consistent role assignments and handoff patterns
  • Existing agent composition used often, should be formalized
  • Complex process naturally breaks into phases or specialties handled by different agents
  • Want to define coordinated group for sprint-based, pipeline-based, or parallel work

Inputs

  • Required: Team name (lowercase kebab-case, e.g., data-pipeline-review)
  • Required: Team purpose (one paragraph describing what problem needs many agents)
  • Required: Lead agent (must exist in agents/_registry.yml)
  • Optional: Coordination pattern (default: hub-and-spoke). One of: hub-and-spoke, sequential, parallel, timeboxed, adaptive
  • Optional: Number of members (default: 3-4; recommended range: 2-5)
  • Optional: Source material (existing workflow, runbook, or ad-hoc team composition to formalize)

Steps

Step 1: Define Team Purpose

Spell out what problem needs many agents working together. Valid team purpose must answer:

  1. What outcome does this team deliver? (e.g., comprehensive review report, deployed application, sprint increment)
  2. Why can't single agent do this? Identify at least two distinct specialties or perspectives needed.
  3. When should this team activate? Define trigger conditions.

Write purpose as one paragraph human or agent can read to decide whether to activate this team.

Got: Clear paragraph explaining team's value. At least two distinct specialties identified.

If fail: Cannot identify two distinct specialties? Task likely does not need team. Use single agent with multiple skills instead.

Step 2: Select Lead Agent

Lead agent orchestrates team. Pick agent from agents/_registry.yml that:

  • Has domain expertise relevant to team's primary output
  • Can decompose incoming requests into subtasks for other members
  • Can synthesize results from many reviewers into coherent deliverable
# List all available agents
grep "^  - id:" agents/_registry.yml

Lead must also show as member in team composition (lead always a member).

Got: One agent picked as lead. Confirmed to exist in agents registry.

If fail: No existing agent fits lead role? Create one first using create-agent skill (or agents/_template.md by hand). Do not create team with lead that does not exist as agent definition.

Step 3: Select Member Agents

Pick 2-5 members (including lead) with clear, non-overlapping responsibilities. For each member, define:

  • id: Agent name from agents registry
  • role: Short title (e.g., "Quality Reviewer", "Security Auditor", "Architecture Reviewer")
  • responsibilities: One sentence describing what this member does that no other member does
# Verify each candidate agent exists
grep "id: agent-name-here" agents/_registry.yml

Validate non-overlap: no two members should have same primary responsibility. Responsibilities overlap? Merge roles or sharpen boundaries.

Got: 2-5 members picked, each with unique role and clear responsibilities. All confirmed in agents registry.

If fail: Needed agent does not exist? Create first. Responsibilities overlap between two members? Rewrite to clarify boundaries or drop one member.

Step 4: Choose Coordination Pattern

Pick pattern fitting team's workflow. Five patterns and use cases:

PatternWhen to UseExample Teams
hub-and-spokeLead distributes tasks, collects results, synthesizes. Best for review and audit workflows.r-package-review, gxp-compliance-validation, ml-data-science-review
sequentialEach agent builds on prior agent's output. Best for pipelines and staged workflows.fullstack-web-dev, tending
parallelAll agents work at once on independent subtasks. Best when subtasks have no dependencies.devops-platform-engineering
timeboxedWork organized into fixed-length iterations. Best for ongoing project work with backlog.scrum-team
adaptiveTeam self-organizes based on task. Best for unknown or highly variable tasks.opaque-team

Decision guide:

  • Lead must see all results before producing output → hub-and-spoke
  • Agent B needs agent A's output to start → sequential
  • All agents can work without seeing each other's output → parallel
  • Work spans many iterations with planning ceremonies → timeboxed
  • Cannot predict task structure in advance → adaptive

Got: One coordination pattern picked with clear rationale.

If fail: Unsure? Default to hub-and-spoke. Most common pattern, works for most review and analysis workflows.

Step 5: Design Task Decomposition

Define how typical incoming request splits across team members. Structure as phases:

  1. Setup phase: What lead does to analyze request and create tasks
  2. Execution phase: What each member works on (in parallel, in sequence, or per-sprint depending on coordination pattern)
  3. Synthesis phase: How results collected and final deliverable produced

For each member, list 3-5 concrete tasks they would do on typical request. These tasks show in both "Task Decomposition" prose section and CONFIG block's tasks list.

Got: Phase-structured decomposition with concrete tasks per member, matching picked coordination pattern.

If fail: Tasks too vague (e.g., "reviews things")? Make specific (e.g., "reviews code style against tidyverse style guide, checks test coverage, evaluates error message quality").

Step 6: Write the Team File

Copy template. Fill in all sections:

cp teams/_template.md teams/<team-name>.md

Fill in these sections in order:

  1. YAML frontmatter: name, description, lead, version ("1.0.0"), author, created, updated, tags, coordination, members[] (each with id, role, responsibilities)
  2. Title: # Team Name (human-readable, title case)
  3. Introduction: One paragraph summary
  4. Purpose: Why this team exists, what specialties it combines
  5. Team Composition: Table with Member, Agent, Role, Focus Areas columns
  6. Coordination Pattern: Prose description plus ASCII diagram of flow
  7. Task Decomposition: Phased breakdown with concrete tasks per member
  8. Configuration: Machine-readable CONFIG block (see Step 7)
  9. Usage Scenarios: 2-3 concrete scenarios with example user prompts
  10. Limitations: 3-5 known constraints
  11. See Also: Links to member agent files and related skills/teams

Got: Complete team file with all sections filled in. No placeholder text left from template.

If fail: Compare against existing team file (e.g., teams/r-package-review.md) to verify structure. Search for template placeholder strings like "your-team-name" or "another-agent" to find unfilled sections.

Step 7: Write the CONFIG Block

CONFIG block between <!-- CONFIG:START --> and <!-- CONFIG:END --> markers gives machine-readable YAML for tooling. Structure as follows:

<!-- CONFIG:START -->
```yaml
team:
  name: <team-name>
  lead: <lead-agent-id>
  coordination: <pattern>
  members:
    - agent: <agent-id>
      role: <role-title>
      subagent_type: <agent-id>  # Claude Code subagent type for spawning
    # ... repeat for each member
  tasks:
    - name: <task-name>
      assignee: <agent-id>
      description: <one-line description>
    # ... repeat for each task
    - name: synthesize-report  # final task if hub-and-spoke
      assignee: <lead-agent-id>
      description: <synthesis description>
      blocked_by: [<prior-task-names>]  # for dependency ordering
```
<!-- CONFIG:END -->

subagent_type field maps to Claude Code agent types. For agents defined in .claude/agents/, use agent id as subagent_type. Use blocked_by for task dependencies (e.g., synthesis blocked by all review tasks).

Got: CONFIG block is valid YAML. All agents match those in frontmatter members list. Task dependencies form valid DAG (no cycles).

If fail: Validate YAML syntax. Verify every assignee in tasks list matches agent in members list. Check blocked_by references only task names defined earlier in list.

Step 8: Add to Registry

Edit teams/_registry.yml. Add new team:

- id: <team-name>
  path: <team-name>.md
  lead: <lead-agent-id>
  members: [<agent-id-1>, <agent-id-2>, ...]
  coordination: <pattern>
  description: <one-line description matching frontmatter>

Bump total_teams count at top of registry (currently 8; becomes 9 after adding one team).

# Verify the entry was added
grep "id: <team-name>" teams/_registry.yml

Got: New entry shows in registry. total_teams count incremented by one.

If fail: Team name already in registry? Pick different name or update existing entry. Verify YAML indentation matches existing entries.

Step 9: Run README Automation

Regenerate README files from updated registry:

npm run update-readmes

Updates dynamic sections in teams/README.md and other files with <!-- AUTO:START --> / <!-- AUTO:END --> markers referencing team data.

Got: Command exits 0. teams/README.md now lists new team.

If fail: Run npm run check-readmes to see which files out of sync. Script fails? Verify package.json exists in repo root and js-yaml installed (npm install).

Step 10: Verify Team Activation

Test team can be activated in Claude Code:

User: Use the <team-name> team to <typical task description>

Claude reads teams/<team-name>.md, extracts CONFIG block, orchestrates activation:

  1. Calls TeamCreate with team name and description
  2. Spawns teammates via Agent tool using each member's subagent_type from CONFIG block
  3. Creates tasks via TaskCreate with blocked_by dependencies from CONFIG block
  4. Lead agent coordinates work following coordination pattern

Note: Teams not auto-discovered from .claude/teams/. Claude reads definition directly from teams/ when asked.

Got: Claude reads team file, creates team via TeamCreate, spawns right agents, follows coordination pattern.

If fail: Verify team file at teams/<team-name>.md (not in subdirectory). Check all member agents exist in agents/. Confirm CONFIG block has valid YAML with subagent_type for each member. Confirm team listed in teams/_registry.yml.

Step 11: Scaffold Translations

Required for all teams. This step applies to both human authors and AI agents following this procedure. Do not skip — missing translations pile into stale backlog.

Scaffold translation files for all 4 supported locales right after committing new team:

for locale in de zh-CN ja es; do
  npm run translate:scaffold -- teams <team-name> "$locale"
done

Then translate scaffolded prose in each file (code blocks and IDs stay English). Finally regenerate status files:

npm run translation:status

Got: 4 files created at i18n/{de,zh-CN,ja,es}/teams/<team-name>.md, all with source_commit matching current HEAD. npm run validate:translations shows 0 stale warnings for new team.

If fail: Scaffold fails? Verify team exists in teams/_registry.yml. Status files don't update? Run npm run translation:status explicitly.

Checks

  • Team file exists at teams/<team-name>.md
  • YAML frontmatter parses without errors
  • All required frontmatter fields present: name, description, lead, version, author, coordination, members[]
  • Each member in frontmatter has id, role, responsibilities
  • All sections present: Purpose, Team Composition, Coordination Pattern, Task Decomposition, Configuration, Usage Scenarios, Limitations, See Also
  • CONFIG block exists between <!-- CONFIG:START --> and <!-- CONFIG:END --> markers
  • CONFIG block YAML valid and parseable
  • All member agent ids exist in agents/_registry.yml
  • Lead agent shows in members list
  • No two members share same primary responsibility
  • Team listed in teams/_registry.yml with right path, lead, members, coordination
  • total_teams count in registry incremented
  • npm run update-readmes finishes without errors

Pitfalls

  • Too many members: Teams with more than 5 members → hard to coordinate. Overhead of distributing tasks and synthesizing results outweighs benefit of extra perspectives. Split into two teams or cut to essential specialties.
  • Overlapping responsibilities: Two members both "review code quality"? Findings will conflict, lead wastes time deduplicating. Each member must have clearly distinct focus area.
  • Wrong coordination pattern: Using hub-and-spoke when agents need each other's output (should be sequential), or using sequential when agents can work independently (should be parallel). Review decision guide in Step 4.
  • Missing CONFIG block: CONFIG block not optional prose decoration. Machine-readable spec Claude uses to orchestrate TeamCreate, agent spawning, task creation. Without it, team only activatable through ad-hoc prose interpretation, less reliable.
  • Lead agent not in members list: Lead must also show as member with own role and responsibilities. Lead who only "coordinates" without substantive work wastes slot. Give lead concrete review or synthesis responsibility.

See Also

  • create-skill - follows same meta-pattern for creating SKILL.md files
  • create-agent - create agent definitions serving as team members
  • commit-changes - commit new team file and registry updates

GitHub 仓库

pjt222/agent-almanac
路径: i18n/caveman/skills/create-team
0
agentsagentskillsai-assisted-developmentclaude-codeskillsteams

相关推荐技能

content-collections

Content Collections 是一个 TypeScript 优先的构建工具,可将本地 Markdown/MDX 文件转换为类型安全的数据集合。它专为构建博客、文档站和内容密集型 Vite+React 应用而设计,提供基于 Zod 的自动模式验证。该工具涵盖从 Vite 插件配置、MDX 编译到生产环境部署的完整工作流。

查看技能

polymarket

这个Claude Skill为开发者提供完整的Polymarket预测市场开发支持,涵盖API调用、交易执行和市场数据分析。关键特性包括实时WebSocket数据流,可监控实时交易、订单和市场动态。开发者可用它构建预测市场应用、实施交易策略并集成实时市场预测功能。

查看技能

creating-opencode-plugins

该Skill帮助开发者创建OpenCode插件,用于接入命令、文件、LSP等25+种事件。它提供了插件结构、事件API规范和JavaScript/TypeScript实现模式,适合需要拦截操作、扩展功能或自定义事件处理的场景。开发者可通过它快速构建响应式模块来增强OpenCode AI助手的能力。

查看技能

sglang

SGLang是一个专为LLM设计的高性能推理框架,特别适用于需要结构化输出的场景。它通过RadixAttention前缀缓存技术,在处理JSON、正则表达式、工具调用等具有重复前缀的复杂工作流时,能实现极速生成。如果你正在构建智能体或多轮对话系统,并追求远超vLLM的推理性能,SGLang是理想选择。

查看技能