[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M6CQW147GA64QQA3HT7MCW4D":3,"skill-same:01M6CQW147GA64QQA3HT7MCW4D":71},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":16,"categories":17,"keywords":23,"publishedAt":33,"linkOnly":34,"markdownContent":35,"markdownBody":36,"frontMatter":37,"refName":38,"commitSha":39,"skillDirectory":40,"forksCount":14,"repositoryHtmlUrl":41,"repositoryHomepageUrl":6,"ownerHtmlUrl":42,"ownerBlogUrl":43,"ownerDescription":6,"updatedAt":44,"install":45,"resources":58},"01M6CQW147GA64QQA3HT7MCW4D","agent-teams",null,"Orchestrate Claude Code agent teams for parallel multi-agent collaboration","dundas\u002Fagentbootup-public","dundas","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F627426?v=4","01M6CQVP2KHXEP7JH9WKKF59S7","David K. Dundas","01M6CQVP76D48NCMMXYDK9CBA8",0,"MIT","https:\u002F\u002Fgithub.com\u002Fdundas\u002Fagentbootup-public\u002Fblob\u002Fmain\u002Ftemplates\u002F.claude\u002Fskills\u002Fagent-teams\u002FSKILL.md",[18],{"categoryNo":19,"categoryName":20,"taxonomyType":21,"taxonomyTypeDesc":22},"01JBSEEDCAT000000000000023","Skill与智能体开发","SCENARIO","使用场景",[24,25,26,27,28,29,30,31,32],"teams","agent teams","orchestration","collaboration","agent collaboration","code","orchestrate","multi-agent","parallel","2026-11-10 10:12:30",false,"---\nname: agent-teams\ndescription: Orchestrate Claude Code agent teams for parallel multi-agent collaboration\ncategory: orchestration\n---\n\n# Agent Teams Skill\n\n## Purpose\nGuide effective use of Claude Code's agent teams feature for parallel multi-agent collaboration with best practices, templates, and orchestration patterns.\n\n## Prerequisites\n\n**Enable agent teams** in `~\u002F.claude\u002Fsettings.json`:\n```json\n{\n  \"env\": {\n    \"CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS\": \"1\"\n  }\n}\n```\n\n## When to Use Agent Teams\n\n### ✅ Good Use Cases\n\n**Research and Review** (Best ROI):\n- Multiple teammates investigate different aspects simultaneously\n- Teammates challenge each other's findings\n- Example: \"Spawn 3 reviewers - security, performance, test coverage\"\n\n**New Features** (Parallel Development):\n- Each teammate owns separate layer\u002Fmodule\n- Minimal file overlap\n- Example: \"Spawn frontend, backend, and security reviewer teammates\"\n\n**Debugging with Competing Hypotheses**:\n- Teammates test different theories in parallel\n- Converge on answer faster through debate\n- Example: \"Spawn 5 teammates to investigate different root causes\"\n\n**Cross-Layer Coordination**:\n- Changes span frontend, backend, tests\n- Each teammate owns different concern\n- Example: \"Spawn API designer, implementer, and test engineer\"\n\n### ❌ Poor Use Cases\n\n- Sequential tasks (use single session)\n- Same-file edits (conflict risk - use single session)\n- Simple tasks (overhead not worth it - use subagents)\n- Work with many dependencies (coordination overhead)\n\n## Team Templates\n\n### 1. Code Review Team (3 reviewers)\n\n```\nCreate an agent team to review PR #{PR_NUMBER}. Spawn three reviewers:\n\n1. **Security Reviewer** (Opus for critical thinking)\n   - Focus: Security vulnerabilities, auth, input validation\n   - Check: SQL injection, XSS, CSRF, secrets exposure\n   - Severity: Rate findings as critical\u002Fhigh\u002Fmedium\u002Flow\n\n2. **Performance Reviewer** (Sonnet)\n   - Focus: Performance implications, scalability\n   - Check: N+1 queries, memory leaks, algorithm complexity\n   - Measure: Impact on latency, throughput, resource usage\n\n3. **Test Coverage Reviewer** (Sonnet)\n   - Focus: Test adequacy, edge cases\n   - Check: Missing tests, untested paths, flaky tests\n   - Ensure: All new code has tests, critical paths covered\n\nHave each reviewer:\n- Review independently\n- Share findings with other reviewers\n- Challenge assumptions\n- Update shared findings document\n\nAfter all reviews complete, synthesize into single report with:\n- Critical blockers (must fix before merge)\n- Important issues (should fix)\n- Suggestions (nice to have)\n```\n\n### 2. Feature Development Team (3 specialists)\n\n```\nCreate an agent team to implement [FEATURE_NAME]. Use delegate mode.\n\n1. **Backend Engineer** (Sonnet)\n   - Owns: API endpoints, business logic, database\n   - Files: src\u002Fapi\u002F*, src\u002Fservices\u002F*, migrations\u002F*\n   - Delivers: Tested API with docs\n\n2. **Frontend Engineer** (Sonnet)\n   - Owns: UI components, state management, routing\n   - Files: src\u002Fcomponents\u002F*, src\u002Fpages\u002F*, src\u002Fhooks\u002F*\n   - Delivers: Tested components with Storybook\n\n3. **Security Reviewer** (Opus)\n   - Monitors: Both backend and frontend work\n   - Reviews: Each implementation before merge\n   - Blocks: Insecure patterns, missing validation\n\nWorkflow:\n1. Backend implements API first\n2. Security reviews backend (blocks if issues)\n3. Frontend builds UI against API\n4. Security reviews frontend (blocks if issues)\n5. Integration testing\n6. Synthesize PR description\n\nTask dependencies:\n- Frontend blocked by backend completion\n- All blocked by security approval\n```\n\n### 3. Research Team (5 investigators)\n\n```\nResearch [TOPIC] using agent team with competing perspectives:\n\n1. **Optimist** - Find why this is the best approach\n2. **Pessimist** - Find why this won't work\n3. **Pragmatist** - Find middle ground, real-world constraints\n4. **Architect** - Evaluate technical feasibility, integration\n5. **Devil's Advocate** - Challenge all assumptions, find edge cases\n\nProcess:\n1. Each teammate researches independently (30 min)\n2. Share findings via broadcast\n3. Debate findings - try to disprove each other\n4. Iterate: respond to challenges with evidence\n5. Converge on consensus (or documented disagreement)\n\nOutput: Research document with:\n- Consensus findings (what everyone agrees on)\n- Disagreements (what's still unclear)\n- Recommendations (with confidence levels)\n- Action items (what to build\u002Ftest next)\n```\n\n### 4. Bug Investigation Team (4 theories)\n\n```\nInvestigate [BUG_DESCRIPTION] with parallel hypothesis testing:\n\nSpawn 4 teammates, each investigating a different theory:\n\n1. **Frontend Theory** - \"Bug is in client-side state management\"\n   - Check: React state, Redux, local storage\n   - Test: Reproduce with different browsers\n\n2. **Backend Theory** - \"Bug is in API response handling\"\n   - Check: API logs, database queries, error handling\n   - Test: Reproduce with API mocking\n\n3. **Race Condition Theory** - \"Bug is timing-dependent\"\n   - Check: Async operations, promises, event handlers\n   - Test: Reproduce with delays, fast\u002Fslow networks\n\n4. **Data Theory** - \"Bug is specific data triggering edge case\"\n   - Check: Data validation, edge cases, null handling\n   - Test: Reproduce with different data samples\n\nProcess:\n1. Each teammate investigates their theory independently\n2. Share findings every 15 minutes\n3. **Actively try to disprove other theories**\n4. Theory that survives scrutiny is likely root cause\n5. Once consensus reached, implement fix\n\nOutput: Root cause analysis with evidence + fix PR\n```\n\n### 5. Refactoring Team (3 layers)\n\n```\nRefactor [MODULE_NAME] using layered team approach:\n\n1. **Architect Teammate** (Opus) - Plan approval required\n   - Design: New architecture, interfaces, data flow\n   - Output: ADR (Architecture Decision Record)\n   - Waits: For lead approval before implementation\n\n2. **Implementer Teammate** (Sonnet)\n   - Implements: Architecture from ADR\n   - Follows: Patterns from ADR exactly\n   - Tests: Each component as built\n\n3. **Migration Teammate** (Sonnet)\n   - Migrates: Existing code to new architecture\n   - Ensures: Backward compatibility\n   - Validates: All existing tests pass\n\nWorkflow:\n1. Architect designs (in plan mode)\n2. Lead reviews and approves design\n3. Implementer builds new architecture\n4. Migration teammate updates existing code\n5. Both run full test suite\n6. Lead synthesizes migration guide\n```\n\n## Best Practices\n\n### Task Sizing\n\n**Good task size**: 30-60 minutes per task\n- ✅ Self-contained unit (function, file, review)\n- ✅ Clear deliverable\n- ✅ Minimal dependencies\n\n**If lead not creating enough tasks**:\n```\nSplit the work into smaller pieces. Aim for 5-6 tasks per teammate.\n```\n\n### Context Provision\n\n**Bad** (too vague):\n```\nSpawn a security reviewer\n```\n\n**Good** (specific context):\n```\nSpawn a security reviewer with the prompt: \"Review src\u002Fauth\u002F for\nsecurity vulnerabilities. Focus on token handling and session management.\nApp uses JWT in httpOnly cookies. Check for: timing attacks, token leakage,\nsession fixation, CSRF. Rate findings as critical\u002Fhigh\u002Fmedium\u002Flow.\"\n```\n\n### File Ownership\n\n**Prevent conflicts** by assigning file ownership:\n```\nBackend teammate owns: src\u002Fapi\u002F*, src\u002Fservices\u002F*\nFrontend teammate owns: src\u002Fcomponents\u002F*, src\u002Fpages\u002F*\nSecurity teammate: read-only review, no file edits\n```\n\n### Monitor Progress\n\n**Check in regularly**:\n```\nWhat's the status of each teammate? Show me their progress.\n```\n\n**Redirect if stuck**:\n```\nTell the frontend teammate to use the existing Button component\ninstead of creating a new one.\n```\n\n### Lead Management\n\n**If lead starts implementing instead of delegating**:\n```\nWait for your teammates to complete their tasks before proceeding.\nFocus on coordination only.\n```\n\n**Use delegate mode** (Shift+Tab) to prevent lead from coding.\n\n## Common Patterns\n\n### Pattern 1: Adversarial Review\n\n```\nCreate agent team to review [DESIGN_DOC]:\n- Spawn 2 reviewers with opposing mandates\n- Reviewer 1: Find why this design is brilliant\n- Reviewer 2: Find why this design will fail\n- Have them debate each point\n- Lead synthesizes balanced perspective\n```\n\n### Pattern 2: Parallel Exploration\n\n```\nCreate agent team to explore [N] different approaches to [PROBLEM]:\n- Spawn N teammates (3-5 optimal)\n- Each teammate prototypes different approach\n- All share findings after 30 minutes\n- Team votes on best approach\n- Lead synthesizes decision rationale\n```\n\n### Pattern 3: Layered Implementation\n\n```\nCreate agent team with staged approval:\n1. Architect (plan approval required) → designs\n2. Lead reviews and approves design\n3. Implementer (Sonnet) → builds\n4. Tester (Sonnet) → validates\n5. Lead synthesizes docs\n```\n\n### Pattern 4: Swarm Intelligence\n\n```\nCreate agent team to solve [HARD_PROBLEM]:\n- Spawn 5 teammates with diverse approaches\n- Each shares progress every 10 minutes\n- Teammates borrow successful strategies from others\n- Evolves toward optimal solution through iteration\n- Lead tracks convergence\n```\n\n## Troubleshooting\n\n### Teammates Not Spawning\n\n**Check if task is complex enough**:\n- Agent teams only for parallel-suitable work\n- For simple tasks, Claude uses single session or subagents\n\n**Explicitly request team**:\n```\nCreate an agent team for this. I want to see parallel work.\n```\n\n### Too Many Permission Prompts\n\n**Pre-approve in `~\u002F.claude\u002Fpermissions.json`**:\n```json\n{\n  \"allow\": {\n    \"Write\": [\"src\u002F**\u002F*\"],\n    \"Edit\": [\"src\u002F**\u002F*\"],\n    \"Bash\": [\"npm test\", \"npm run build\"]\n  }\n}\n```\n\n### Teammates Conflicting\n\n**Assign clear file ownership**:\n```\nBackend owns src\u002Fapi\u002F\nFrontend owns src\u002Fcomponents\u002F\nNo overlap allowed.\n```\n\n### Team Runs Too Long\n\n**Set time boundaries**:\n```\nEach teammate has 30 minutes to complete their investigation.\nReport findings even if incomplete.\n```\n\n### Lead Doing Work Instead of Delegating\n\n**Enable delegate mode**:\n```\nPress Shift+Tab to enable delegate mode\n```\n\nOr remind:\n```\nStop implementing. Your job is to coordinate teammates only.\n```\n\n## Architecture\n\n### Storage Locations\n\n- **Team config**: `~\u002F.claude\u002Fteams\u002F{team-name}\u002Fconfig.json`\n- **Task list**: `~\u002F.claude\u002Ftasks\u002F{team-name}\u002F`\n- **Teammate members**: Listed in config.json\n\n### Communication Methods\n\n**Direct message** (one-to-one):\n```\nmessage({recipientId: \"backend-engineer\", message: \"...\"})\n```\n\n**Broadcast** (one-to-all):\n```\nbroadcast({message: \"All teammates: share your progress\"})\n```\n\n**Shared task list** (implicit coordination):\n- Teammates see all tasks\n- Can claim available work\n- File locking prevents conflicts\n\n## Token Cost Management\n\n**Agent teams use ~3-5x tokens of single session**\n\n### Minimize Cost\n\n1. **Use Haiku for simple teammates**:\n   ```\n   Spawn test runner teammate with Haiku model\n   ```\n\n2. **Start small, scale up**:\n   - Begin with 2-3 teammates\n   - Add more only if needed\n\n3. **Time-box work**:\n   ```\n   Each teammate has 30 minutes max\n   ```\n\n4. **Avoid broadcast overuse**:\n   - Broadcast = message to ALL teammates\n   - Costs scale with team size\n   - Use direct messages when possible\n\n### When Cost is Worth It\n\n- ✅ Critical security review (parallel expertise)\n- ✅ Complex feature (parallel implementation)\n- ✅ Research (diverse perspectives)\n- ❌ Simple bug fix (use single session)\n- ❌ Documentation update (use single session)\n\n## Examples\n\n### Example 1: Security-First Feature Development\n\n```\nI'm adding payment processing. Create an agent team with security review at each step:\n\n1. Backend Engineer (Sonnet)\n   - Implement payment API\n   - Use Stripe SDK\n   - Store minimal PII\n\n2. Frontend Engineer (Sonnet)\n   - Build payment form\n   - Use Stripe Elements (no card data on our server)\n   - Handle errors gracefully\n\n3. Security Reviewer (Opus)\n   - Review each implementation before allowing next step\n   - Check: PCI DSS compliance, no card data logged, HTTPS only\n   - Block: Any violations\n\nUse delegate mode. Backend goes first, must pass security review before frontend starts.\n```\n\n### Example 2: Adversarial Design Review\n\n```\nReview this architecture design document: [DOC_URL]\n\nCreate agent team with adversarial approach:\n\n1. **Advocate** - Find strengths, why this design is optimal\n2. **Critic** - Find weaknesses, why this design will fail\n3. **Moderator** - Synthesize debate, propose improvements\n\nProcess:\n- Advocate and Critic read doc independently\n- Each presents case (5 minutes)\n- Debate each point (20 minutes)\n- Moderator synthesizes balanced view\n- Lead creates updated design with improvements\n```\n\n### Example 3: Parallel Hypothesis Testing\n\n```\nOur API is slow. Create agent team to test different optimization theories:\n\n1. **Database Theory** - \"Queries are slow\"\n   - Profile queries with EXPLAIN ANALYZE\n   - Test: Add indexes, check improvement\n\n2. **Network Theory** - \"Network latency is high\"\n   - Profile with network inspector\n   - Test: Add CDN, check improvement\n\n3. **Code Theory** - \"Algorithm is inefficient\"\n   - Profile with performance tools\n   - Test: Optimize hot paths, check improvement\n\n4. **Scale Theory** - \"System is overloaded\"\n   - Check: CPU, memory, disk I\u002FO\n   - Test: Vertical scaling, check improvement\n\nEach teammate:\n- Measures baseline performance\n- Implements their optimization\n- Measures improvement\n- Shares results with team\n\nWinner: Theory with biggest measurable improvement\n```\n\n## Integration with Other Skills\n\n**Combine with**:\n- **brain-briefing**: Get context before spawning team\n- **task-processor**: Break PRD into tasks for team\n- **pr-review-loop**: Use team for multi-reviewer approval\n\n## Limitations (Experimental)\n\n- No session resumption with in-process teammates\n- Task status can lag (manually update if needed)\n- Shutdown can be slow (teammates finish current work)\n- One team per session (clean up before new team)\n- No nested teams (teammates can't spawn teams)\n- Lead is fixed (can't promote teammate)\n\n## Quick Reference\n\n### Start Team\n```\nCreate an agent team for [TASK]. Spawn [N] teammates: [ROLES]\n```\n\n### Talk to Teammate\n```\nShift+Up\u002FDown to select, then type message\n```\n\n### Check Progress\n```\nWhat's the status of each teammate?\n```\n\n### Shut Down Teammate\n```\nAsk [teammate-name] to shut down\n```\n\n### Clean Up\n```\nClean up the team\n```\n\n### Enable Delegate Mode\n```\nPress Shift+Tab\n```\n\n---\n\n**Created**: 2026-02-06\n**Requires**: `CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1`\n**Documentation**: `.claude\u002Fdocs\u002FAGENT_TEAMS.md`\n","# Agent Teams Skill\n\n## Purpose\nGuide effective use of Claude Code's agent teams feature for parallel multi-agent collaboration with best practices, templates, and orchestration patterns.\n\n## Prerequisites\n\n**Enable agent teams** in `~\u002F.claude\u002Fsettings.json`:\n```json\n{\n  \"env\": {\n    \"CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS\": \"1\"\n  }\n}\n```\n\n## When to Use Agent Teams\n\n### ✅ Good Use Cases\n\n**Research and Review** (Best ROI):\n- Multiple teammates investigate different aspects simultaneously\n- Teammates challenge each other's findings\n- Example: \"Spawn 3 reviewers - security, performance, test coverage\"\n\n**New Features** (Parallel Development):\n- Each teammate owns separate layer\u002Fmodule\n- Minimal file overlap\n- Example: \"Spawn frontend, backend, and security reviewer teammates\"\n\n**Debugging with Competing Hypotheses**:\n- Teammates test different theories in parallel\n- Converge on answer faster through debate\n- Example: \"Spawn 5 teammates to investigate different root causes\"\n\n**Cross-Layer Coordination**:\n- Changes span frontend, backend, tests\n- Each teammate owns different concern\n- Example: \"Spawn API designer, implementer, and test engineer\"\n\n### ❌ Poor Use Cases\n\n- Sequential tasks (use single session)\n- Same-file edits (conflict risk - use single session)\n- Simple tasks (overhead not worth it - use subagents)\n- Work with many dependencies (coordination overhead)\n\n## Team Templates\n\n### 1. Code Review Team (3 reviewers)\n\n```\nCreate an agent team to review PR #{PR_NUMBER}. Spawn three reviewers:\n\n1. **Security Reviewer** (Opus for critical thinking)\n   - Focus: Security vulnerabilities, auth, input validation\n   - Check: SQL injection, XSS, CSRF, secrets exposure\n   - Severity: Rate findings as critical\u002Fhigh\u002Fmedium\u002Flow\n\n2. **Performance Reviewer** (Sonnet)\n   - Focus: Performance implications, scalability\n   - Check: N+1 queries, memory leaks, algorithm complexity\n   - Measure: Impact on latency, throughput, resource usage\n\n3. **Test Coverage Reviewer** (Sonnet)\n   - Focus: Test adequacy, edge cases\n   - Check: Missing tests, untested paths, flaky tests\n   - Ensure: All new code has tests, critical paths covered\n\nHave each reviewer:\n- Review independently\n- Share findings with other reviewers\n- Challenge assumptions\n- Update shared findings document\n\nAfter all reviews complete, synthesize into single report with:\n- Critical blockers (must fix before merge)\n- Important issues (should fix)\n- Suggestions (nice to have)\n```\n\n### 2. Feature Development Team (3 specialists)\n\n```\nCreate an agent team to implement [FEATURE_NAME]. Use delegate mode.\n\n1. **Backend Engineer** (Sonnet)\n   - Owns: API endpoints, business logic, database\n   - Files: src\u002Fapi\u002F*, src\u002Fservices\u002F*, migrations\u002F*\n   - Delivers: Tested API with docs\n\n2. **Frontend Engineer** (Sonnet)\n   - Owns: UI components, state management, routing\n   - Files: src\u002Fcomponents\u002F*, src\u002Fpages\u002F*, src\u002Fhooks\u002F*\n   - Delivers: Tested components with Storybook\n\n3. **Security Reviewer** (Opus)\n   - Monitors: Both backend and frontend work\n   - Reviews: Each implementation before merge\n   - Blocks: Insecure patterns, missing validation\n\nWorkflow:\n1. Backend implements API first\n2. Security reviews backend (blocks if issues)\n3. Frontend builds UI against API\n4. Security reviews frontend (blocks if issues)\n5. Integration testing\n6. Synthesize PR description\n\nTask dependencies:\n- Frontend blocked by backend completion\n- All blocked by security approval\n```\n\n### 3. Research Team (5 investigators)\n\n```\nResearch [TOPIC] using agent team with competing perspectives:\n\n1. **Optimist** - Find why this is the best approach\n2. **Pessimist** - Find why this won't work\n3. **Pragmatist** - Find middle ground, real-world constraints\n4. **Architect** - Evaluate technical feasibility, integration\n5. **Devil's Advocate** - Challenge all assumptions, find edge cases\n\nProcess:\n1. Each teammate researches independently (30 min)\n2. Share findings via broadcast\n3. Debate findings - try to disprove each other\n4. Iterate: respond to challenges with evidence\n5. Converge on consensus (or documented disagreement)\n\nOutput: Research document with:\n- Consensus findings (what everyone agrees on)\n- Disagreements (what's still unclear)\n- Recommendations (with confidence levels)\n- Action items (what to build\u002Ftest next)\n```\n\n### 4. Bug Investigation Team (4 theories)\n\n```\nInvestigate [BUG_DESCRIPTION] with parallel hypothesis testing:\n\nSpawn 4 teammates, each investigating a different theory:\n\n1. **Frontend Theory** - \"Bug is in client-side state management\"\n   - Check: React state, Redux, local storage\n   - Test: Reproduce with different browsers\n\n2. **Backend Theory** - \"Bug is in API response handling\"\n   - Check: API logs, database queries, error handling\n   - Test: Reproduce with API mocking\n\n3. **Race Condition Theory** - \"Bug is timing-dependent\"\n   - Check: Async operations, promises, event handlers\n   - Test: Reproduce with delays, fast\u002Fslow networks\n\n4. **Data Theory** - \"Bug is specific data triggering edge case\"\n   - Check: Data validation, edge cases, null handling\n   - Test: Reproduce with different data samples\n\nProcess:\n1. Each teammate investigates their theory independently\n2. Share findings every 15 minutes\n3. **Actively try to disprove other theories**\n4. Theory that survives scrutiny is likely root cause\n5. Once consensus reached, implement fix\n\nOutput: Root cause analysis with evidence + fix PR\n```\n\n### 5. Refactoring Team (3 layers)\n\n```\nRefactor [MODULE_NAME] using layered team approach:\n\n1. **Architect Teammate** (Opus) - Plan approval required\n   - Design: New architecture, interfaces, data flow\n   - Output: ADR (Architecture Decision Record)\n   - Waits: For lead approval before implementation\n\n2. **Implementer Teammate** (Sonnet)\n   - Implements: Architecture from ADR\n   - Follows: Patterns from ADR exactly\n   - Tests: Each component as built\n\n3. **Migration Teammate** (Sonnet)\n   - Migrates: Existing code to new architecture\n   - Ensures: Backward compatibility\n   - Validates: All existing tests pass\n\nWorkflow:\n1. Architect designs (in plan mode)\n2. Lead reviews and approves design\n3. Implementer builds new architecture\n4. Migration teammate updates existing code\n5. Both run full test suite\n6. Lead synthesizes migration guide\n```\n\n## Best Practices\n\n### Task Sizing\n\n**Good task size**: 30-60 minutes per task\n- ✅ Self-contained unit (function, file, review)\n- ✅ Clear deliverable\n- ✅ Minimal dependencies\n\n**If lead not creating enough tasks**:\n```\nSplit the work into smaller pieces. Aim for 5-6 tasks per teammate.\n```\n\n### Context Provision\n\n**Bad** (too vague):\n```\nSpawn a security reviewer\n```\n\n**Good** (specific context):\n```\nSpawn a security reviewer with the prompt: \"Review src\u002Fauth\u002F for\nsecurity vulnerabilities. Focus on token handling and session management.\nApp uses JWT in httpOnly cookies. Check for: timing attacks, token leakage,\nsession fixation, CSRF. Rate findings as critical\u002Fhigh\u002Fmedium\u002Flow.\"\n```\n\n### File Ownership\n\n**Prevent conflicts** by assigning file ownership:\n```\nBackend teammate owns: src\u002Fapi\u002F*, src\u002Fservices\u002F*\nFrontend teammate owns: src\u002Fcomponents\u002F*, src\u002Fpages\u002F*\nSecurity teammate: read-only review, no file edits\n```\n\n### Monitor Progress\n\n**Check in regularly**:\n```\nWhat's the status of each teammate? Show me their progress.\n```\n\n**Redirect if stuck**:\n```\nTell the frontend teammate to use the existing Button component\ninstead of creating a new one.\n```\n\n### Lead Management\n\n**If lead starts implementing instead of delegating**:\n```\nWait for your teammates to complete their tasks before proceeding.\nFocus on coordination only.\n```\n\n**Use delegate mode** (Shift+Tab) to prevent lead from coding.\n\n## Common Patterns\n\n### Pattern 1: Adversarial Review\n\n```\nCreate agent team to review [DESIGN_DOC]:\n- Spawn 2 reviewers with opposing mandates\n- Reviewer 1: Find why this design is brilliant\n- Reviewer 2: Find why this design will fail\n- Have them debate each point\n- Lead synthesizes balanced perspective\n```\n\n### Pattern 2: Parallel Exploration\n\n```\nCreate agent team to explore [N] different approaches to [PROBLEM]:\n- Spawn N teammates (3-5 optimal)\n- Each teammate prototypes different approach\n- All share findings after 30 minutes\n- Team votes on best approach\n- Lead synthesizes decision rationale\n```\n\n### Pattern 3: Layered Implementation\n\n```\nCreate agent team with staged approval:\n1. Architect (plan approval required) → designs\n2. Lead reviews and approves design\n3. Implementer (Sonnet) → builds\n4. Tester (Sonnet) → validates\n5. Lead synthesizes docs\n```\n\n### Pattern 4: Swarm Intelligence\n\n```\nCreate agent team to solve [HARD_PROBLEM]:\n- Spawn 5 teammates with diverse approaches\n- Each shares progress every 10 minutes\n- Teammates borrow successful strategies from others\n- Evolves toward optimal solution through iteration\n- Lead tracks convergence\n```\n\n## Troubleshooting\n\n### Teammates Not Spawning\n\n**Check if task is complex enough**:\n- Agent teams only for parallel-suitable work\n- For simple tasks, Claude uses single session or subagents\n\n**Explicitly request team**:\n```\nCreate an agent team for this. I want to see parallel work.\n```\n\n### Too Many Permission Prompts\n\n**Pre-approve in `~\u002F.claude\u002Fpermissions.json`**:\n```json\n{\n  \"allow\": {\n    \"Write\": [\"src\u002F**\u002F*\"],\n    \"Edit\": [\"src\u002F**\u002F*\"],\n    \"Bash\": [\"npm test\", \"npm run build\"]\n  }\n}\n```\n\n### Teammates Conflicting\n\n**Assign clear file ownership**:\n```\nBackend owns src\u002Fapi\u002F\nFrontend owns src\u002Fcomponents\u002F\nNo overlap allowed.\n```\n\n### Team Runs Too Long\n\n**Set time boundaries**:\n```\nEach teammate has 30 minutes to complete their investigation.\nReport findings even if incomplete.\n```\n\n### Lead Doing Work Instead of Delegating\n\n**Enable delegate mode**:\n```\nPress Shift+Tab to enable delegate mode\n```\n\nOr remind:\n```\nStop implementing. Your job is to coordinate teammates only.\n```\n\n## Architecture\n\n### Storage Locations\n\n- **Team config**: `~\u002F.claude\u002Fteams\u002F{team-name}\u002Fconfig.json`\n- **Task list**: `~\u002F.claude\u002Ftasks\u002F{team-name}\u002F`\n- **Teammate members**: Listed in config.json\n\n### Communication Methods\n\n**Direct message** (one-to-one):\n```\nmessage({recipientId: \"backend-engineer\", message: \"...\"})\n```\n\n**Broadcast** (one-to-all):\n```\nbroadcast({message: \"All teammates: share your progress\"})\n```\n\n**Shared task list** (implicit coordination):\n- Teammates see all tasks\n- Can claim available work\n- File locking prevents conflicts\n\n## Token Cost Management\n\n**Agent teams use ~3-5x tokens of single session**\n\n### Minimize Cost\n\n1. **Use Haiku for simple teammates**:\n   ```\n   Spawn test runner teammate with Haiku model\n   ```\n\n2. **Start small, scale up**:\n   - Begin with 2-3 teammates\n   - Add more only if needed\n\n3. **Time-box work**:\n   ```\n   Each teammate has 30 minutes max\n   ```\n\n4. **Avoid broadcast overuse**:\n   - Broadcast = message to ALL teammates\n   - Costs scale with team size\n   - Use direct messages when possible\n\n### When Cost is Worth It\n\n- ✅ Critical security review (parallel expertise)\n- ✅ Complex feature (parallel implementation)\n- ✅ Research (diverse perspectives)\n- ❌ Simple bug fix (use single session)\n- ❌ Documentation update (use single session)\n\n## Examples\n\n### Example 1: Security-First Feature Development\n\n```\nI'm adding payment processing. Create an agent team with security review at each step:\n\n1. Backend Engineer (Sonnet)\n   - Implement payment API\n   - Use Stripe SDK\n   - Store minimal PII\n\n2. Frontend Engineer (Sonnet)\n   - Build payment form\n   - Use Stripe Elements (no card data on our server)\n   - Handle errors gracefully\n\n3. Security Reviewer (Opus)\n   - Review each implementation before allowing next step\n   - Check: PCI DSS compliance, no card data logged, HTTPS only\n   - Block: Any violations\n\nUse delegate mode. Backend goes first, must pass security review before frontend starts.\n```\n\n### Example 2: Adversarial Design Review\n\n```\nReview this architecture design document: [DOC_URL]\n\nCreate agent team with adversarial approach:\n\n1. **Advocate** - Find strengths, why this design is optimal\n2. **Critic** - Find weaknesses, why this design will fail\n3. **Moderator** - Synthesize debate, propose improvements\n\nProcess:\n- Advocate and Critic read doc independently\n- Each presents case (5 minutes)\n- Debate each point (20 minutes)\n- Moderator synthesizes balanced view\n- Lead creates updated design with improvements\n```\n\n### Example 3: Parallel Hypothesis Testing\n\n```\nOur API is slow. Create agent team to test different optimization theories:\n\n1. **Database Theory** - \"Queries are slow\"\n   - Profile queries with EXPLAIN ANALYZE\n   - Test: Add indexes, check improvement\n\n2. **Network Theory** - \"Network latency is high\"\n   - Profile with network inspector\n   - Test: Add CDN, check improvement\n\n3. **Code Theory** - \"Algorithm is inefficient\"\n   - Profile with performance tools\n   - Test: Optimize hot paths, check improvement\n\n4. **Scale Theory** - \"System is overloaded\"\n   - Check: CPU, memory, disk I\u002FO\n   - Test: Vertical scaling, check improvement\n\nEach teammate:\n- Measures baseline performance\n- Implements their optimization\n- Measures improvement\n- Shares results with team\n\nWinner: Theory with biggest measurable improvement\n```\n\n## Integration with Other Skills\n\n**Combine with**:\n- **brain-briefing**: Get context before spawning team\n- **task-processor**: Break PRD into tasks for team\n- **pr-review-loop**: Use team for multi-reviewer approval\n\n## Limitations (Experimental)\n\n- No session resumption with in-process teammates\n- Task status can lag (manually update if needed)\n- Shutdown can be slow (teammates finish current work)\n- One team per session (clean up before new team)\n- No nested teams (teammates can't spawn teams)\n- Lead is fixed (can't promote teammate)\n\n## Quick Reference\n\n### Start Team\n```\nCreate an agent team for [TASK]. Spawn [N] teammates: [ROLES]\n```\n\n### Talk to Teammate\n```\nShift+Up\u002FDown to select, then type message\n```\n\n### Check Progress\n```\nWhat's the status of each teammate?\n```\n\n### Shut Down Teammate\n```\nAsk [teammate-name] to shut down\n```\n\n### Clean Up\n```\nClean up the team\n```\n\n### Enable Delegate Mode\n```\nPress Shift+Tab\n```\n\n---\n\n**Created**: 2026-02-06\n**Requires**: `CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1`\n**Documentation**: `.claude\u002Fdocs\u002FAGENT_TEAMS.md`\n",{"name":5,"category":26,"description":7},"main","47b89745588c3057f0ccd0517a2eb8967df27054","templates\u002F.claude\u002Fskills\u002Fagent-teams","https:\u002F\u002Fgithub.com\u002Fdundas\u002Fagentbootup-public","https:\u002F\u002Fgithub.com\u002Fdundas","derivative.io","2026-09-02 20:12:15",{"source":8,"skillName":5,"commands":46,"prompt":56,"zipAvailable":57},[47,50,53],{"tool":48,"command":49},"npx","npx skills add dundas\u002Fagentbootup-public --skill agent-teams",{"tool":51,"command":52},"bunx","bunx skills add dundas\u002Fagentbootup-public --skill agent-teams",{"tool":54,"command":55},"pnpm","pnpm dlx skills add dundas\u002Fagentbootup-public --skill agent-teams","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002Fdundas\u002Fagentbootup-public\u002Fblob\u002Fmain\u002Ftemplates\u002F.claude\u002Fskills\u002Fagent-teams\u002FSKILL.md",true,[59,65],{"path":60,"resourceType":61,"resourceTypeDesc":62,"sizeBytes":63,"sourceHtmlUrl":64},"scripts\u002Fcalculate_cac.py","SCRIPT","脚本",3029,"https:\u002F\u002Fgithub.com\u002FMishit18\u002Fclaude-code-templates\u002Fblob\u002Fmain\u002Fcli-tool\u002Fcomponents\u002Fskills\u002Fbusiness-marketing\u002Fmarketing-demand-acquisition\u002Fscripts\u002Fcalculate_cac.py",{"path":66,"resourceType":67,"resourceTypeDesc":68,"sizeBytes":69,"sourceHtmlUrl":70},"skill-bundle-manifest.json","REFERENCE","参考文档",774,"https:\u002F\u002Fgithub.com\u002Fdundas\u002Fagentbootup-public\u002Fblob\u002Fmain\u002Ftemplates\u002F.claude\u002Fskills\u002Fagent-teams\u002Fskill-bundle-manifest.json",[72,82,102,122,131,153],{"skillNo":73,"name":74,"nameZh":6,"description":75,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":76,"categories":77,"keywords":79,"publishedAt":33},"01M6CQZ7DFJB965BJJ5ZZ54CXY","fixture-skill","Secret-free recovery fixture skill.","https:\u002F\u002Fgithub.com\u002Fdundas\u002Fagentbootup-public\u002Fblob\u002Fmain\u002Ftests\u002Fruntime-adapters\u002Ffixtures\u002Fhermes\u002F0.18.2-darwin-arm64-real\u002Froot\u002Fskills\u002Ffixture-skill\u002FSKILL.md",[78],{"categoryNo":19,"categoryName":20,"taxonomyType":21,"taxonomyTypeDesc":22},[80,81],"fixture skill","recovery",{"skillNo":83,"name":84,"nameZh":6,"description":85,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":86,"categories":87,"keywords":93,"publishedAt":33},"01M6CQZ4NN0KVH1Z4E1Y1GJ3X9","user-story-generator","Generate standalone user stories for features without creating a full PRD. Perfect for backlog grooming and story refinement.","https:\u002F\u002Fgithub.com\u002Fdundas\u002Fagentbootup-public\u002Fblob\u002Fmain\u002Ftemplates\u002F.claude\u002Fskills\u002Fuser-story-generator\u002FSKILL.md",[88],{"categoryNo":89,"categoryName":90,"taxonomyType":91,"taxonomyTypeDesc":92},"01JBSEEDCAT000000000000031","产品经理","PROFESSION","适用职业",[94,95,96,97,98,99,100,101],"user story generator","Story","user story","user","prd","documentation","generate","backlog",{"skillNo":103,"name":104,"nameZh":6,"description":105,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":106,"categories":107,"keywords":114,"publishedAt":33},"01M6CQZ16VC27HMWM31CCKK6BC","user-journey-mapper","Map user flows and journeys through your application for UX design and workflow visualization, independent of testing.","https:\u002F\u002Fgithub.com\u002Fdundas\u002Fagentbootup-public\u002Fblob\u002Fmain\u002Ftemplates\u002F.claude\u002Fskills\u002Fuser-journey-mapper\u002FSKILL.md",[108,111],{"categoryNo":109,"categoryName":110,"taxonomyType":21,"taxonomyTypeDesc":22},"01JBSEEDCAT000000000000008","测试与质量",{"categoryNo":112,"categoryName":113,"taxonomyType":91,"taxonomyTypeDesc":92},"01JBSEEDCAT000000000000032","设计师",[115,116,97,99,117,118,119,120,121],"user journey mapper","journey","ux","visualization","workflow","design","testing",{"skillNo":123,"name":124,"nameZh":6,"description":6,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":125,"categories":126,"keywords":127,"publishedAt":33},"01M6CQYXSQQJ4Z5CRAX2SSMHW1","transcript-query","https:\u002F\u002Fgithub.com\u002Fdundas\u002Fagentbootup-public\u002Fblob\u002Fmain\u002Ftemplates\u002F.claude\u002Fskills\u002Ftranscript-query\u002FSKILL.md",[],[128,129,130],"transcript","transcript query","query",{"skillNo":132,"name":133,"nameZh":6,"description":134,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":135,"categories":136,"keywords":141,"publishedAt":33},"01M6CQYT7A1G4CPQTZKXR5TCK9","test-plan-generator","Generate comprehensive E2E test plans from PRDs, task lists, or descriptions with environment-aware prerequisites and issue tracking.","https:\u002F\u002Fgithub.com\u002Fdundas\u002Fagentbootup-public\u002Fblob\u002Fmain\u002Ftemplates\u002F.claude\u002Fskills\u002Ftest-plan-generator\u002FSKILL.md",[137,138],{"categoryNo":109,"categoryName":110,"taxonomyType":21,"taxonomyTypeDesc":22},{"categoryNo":139,"categoryName":140,"taxonomyType":21,"taxonomyTypeDesc":22},"01JBSEEDCAT000000000000010","后端与接口",[142,143,144,145,146,147,148,149,150,151,100,152],"plan","test plan generator","test plan","test","task","issue","e2e test","environment","tracking","quality","e2e",{"skillNo":154,"name":155,"nameZh":6,"description":156,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":157,"categories":158,"keywords":160,"publishedAt":33},"01M6CQYQAWFSPTYE2VTJJVW72G","tasklist-generator","Generate high-level tasks and gated sub-tasks from a PRD, with agent assignments, commit\u002FPR strategy, and relevant files.","https:\u002F\u002Fgithub.com\u002Fdundas\u002Fagentbootup-public\u002Fblob\u002Fmain\u002Ftemplates\u002F.claude\u002Fskills\u002Ftasklist-generator\u002FSKILL.md",[159],{"categoryNo":19,"categoryName":20,"taxonomyType":21,"taxonomyTypeDesc":22},[161,162,163,164,98,100,165,166,167,119],"tasklist generator","gated","files","strategy","tasks","pr","commit"]