메뉴
Use when facing a complex task that benefits from multiple specialized agents working in parallel — spawns, delegates, tracks status, and collects results. Also use when the user says "use agents", "multi-agent", or "parallelize this".
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
SOC 직업 분류 기준
Use when you need to commit and push changes — stages files, commits with co-author tags, pulls with rebase, resolves conflicts, and pushes. Also supports "pr" mode for creating PRs and "clean" mode for pruning merged branches. Also use for "commit", "push", "sync", or "create PR".
Use when the user wants short-and-sweet answers to one or more questions. Restates each question as a bold heading, then answers in 2–5 lines per question. Bullets and small tables allowed when they help. Also use for "quick question", "qq", "answer briefly", or any "/my-qq <question>" invocation.
Use when starting implementation on any project — reads planning docs, dispatches sub-agents, and creates or reiterates docs/acceptance-criteria.md with measurable criteria per feature. Must run before writing implementation code. Also use for "create acceptance criteria", "what are the requirements", or "define done".
Use when starting a new feature or project from a raw idea — asks targeted questions and produces a structured PRD at docs/prd-<slug>.md. Sits upstream of /my-prompt. Also use for "write a PRD", "product requirements", or "what should we build".
Use when code needs rigorous review before shipping — spawns skeptic/believer/referee agents in iterative rounds until findings converge to nitpicks. Use after implementing features, before merging PRs, or when the user asks for a "thorough review", "adversarial review", or "tear this apart".
Use when you want to find and fix gaps in the codebase — dispatches all agent personalities to audit code against CLAUDE.md guardrails and acceptance criteria, documents gaps with fix suggestions, then implements fixes. Also use for "audit the code", "find gaps", or "check against criteria".
| name | my-agent-teams |
| description | Use when facing a complex task that benefits from multiple specialized agents working in parallel — spawns, delegates, tracks status, and collects results. Also use when the user says "use agents", "multi-agent", or "parallelize this". |
| argument-hint | < feature | debug | review | custom > [description] |
| allowed-tools | Read, Glob, Grep, Bash(git:*), Agent |
Spawn coordinated agent teams for tasks too large or complex for a single agent. Each team has a lead who delegates, tracks, and synthesizes.
What: Spin up a coordinated team of agents with a lead, workers, and a reviewer. Tracks progress and synthesizes results. Usage:
/my-agent-teams feature Add webhook signature verification to all endpoints — feature team (architect + implementers + reviewer)/my-agent-teams debug Payment webhook failing silently for Stripe test events — debug team (investigator + reproducer + fixer)/my-agent-teams review src/api/ — review team (code-reviewer + security-reviewer + architect)/my-agent-teams custom 3 agents: one reads all DB schemas, one reads all API routes, one reads all tests — synthesize gaps — custom team
Output: Synthesized report from team lead with individual agent findings merged.| Role | Agent Identity | Model | Task |
|---|---|---|---|
| Lead: Architect | specialists/architect.md | opus | Break feature into tasks, assign to implementers, review integration |
| Worker 1: Implementer | Default | sonnet | Implement task 1 from architect's breakdown |
| Worker 2: Implementer | Default | sonnet | Implement task 2 (parallel with Worker 1) |
| Reviewer | reviewers/code.md | sonnet | Review all implementations against acceptance criteria |
| Role | Agent Identity | Model | Task |
|---|---|---|---|
| Lead: Investigator | debate/skeptic.md | sonnet | Trace the bug — read logs, find root cause, form hypothesis |
| Worker 1: Reproducer | Default | sonnet | Write a minimal reproduction (test case that fails) |
| Worker 2: Fixer | Default | sonnet | Implement fix based on investigator's hypothesis |
| Reviewer | reviewers/code.md | haiku | Verify fix is correct and doesn't introduce regressions |
| Role | Agent Identity | Model | Task |
|---|---|---|---|
| Lead: Code Reviewer | reviewers/code.md | sonnet | Full 10-dimension code review |
| Worker 1: Security | reviewers/security.md | sonnet | Security-focused review with STRIDE |
| Worker 2: Architect | specialists/architect.md | opus | Architecture and design review |
| Synthesizer | debate/referee.md | sonnet | Deduplicate findings, resolve conflicts, final verdict |
User specifies the number of agents and their tasks. Lead is auto-assigned as the first agent.
If the user specifies a template (feature, debug, review), use it. Otherwise parse the custom description to determine:
~/.claude/agents/)Before spawning agents, gather shared context they all need:
Launch all worker agents simultaneously using the Agent tool. Each agent gets:
~/.claude/agents/)Critical: All workers run in parallel. Do NOT wait for one to finish before launching the next.
Match model tier to task complexity. Use the model parameter on the Agent tool:
| Tier | Use for | Cost |
|---|---|---|
| haiku | Verification, formatting, grep searches, simple transforms, scouts (github) | 10-20x cheaper |
| sonnet | Standard implementation, research, coding, code review, debate agents | Default |
| opus | Architecture decisions, strategy, complex judgment, orchestration | Most expensive |
Agent → Tier mapping:
| Agent | Recommended Tier |
|---|---|
| debate/skeptic | sonnet |
| debate/believer | sonnet |
| debate/referee | sonnet |
| reviewers/code | sonnet |
| reviewers/doc | haiku |
| reviewers/security | sonnet |
| scouts/github | haiku |
| specialists/architect | opus |
| specialists/frontend | sonnet |
| specialists/platform | sonnet |
| workers/crawler | haiku |
| Default implementer | sonnet |
| Verification/formatting | haiku |
Example dispatch:
Agent(prompt="...", model="haiku") # for scouts, verification
Agent(prompt="...", model="sonnet") # for implementation, review
Agent(prompt="...", model="opus") # for architecture decisions
Once all agents return:
## Team Report: [task description]
Team: [template name or "custom"] | Agents: [N] | Duration: [time]
### Team Composition
| Role | Identity | Status |
|------|----------|--------|
### Synthesized Findings
[merged, deduplicated results from all agents — sorted by priority]
### Conflicts Resolved
[where agents disagreed and how it was resolved]
### Next Steps
[actionable items from the team's work]
model: haiku for verification/formatting agents to cut costs 10-20x~/.claude/agents/ when a matching personality exists