메뉴
When and how workers use MCP tools during /mission-run. Prefer MCP for live service introspection (Supabase schema, GitHub, etc.) when connections/mcp-registry.md lists the service. Activate in worker subagents during execution.
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
Guidance for picking the right model per role ("droid whispering"). Activate when configuring agents, starting a new mission, or diagnosing why a role is misbehaving. Model strings age fast — verify current names via search per P-2 before relying on the defaults below.
How to set up every external service the project needs WITHOUT asking the user to run commands or follow long tutorials. The orchestrator handles all installation, configuration, MCP setup, and verification. The user provides only the data the orchestrator cannot obtain itself (credentials, choices). MCP servers are preferred over raw API integration where available.
How to run the two-round multiple-choice discovery questionnaire. Defines the 30 default round-1 questions across 6 categories, how to adapt them per project type, and how to generate the 15 round-2 follow-ups from gaps and contradictions.
How to produce the validation contract, draft plan, and tech-decisions from the description and discovery answers. Web-searches for current versions before committing to anything. Targets fine-grained feature decomposition (more features = better). Features here are DRAFTS — the /mission-tasks phase enriches them. Activate during /mission-plan.
Mandatory handoff template every worker fills out before exiting. The pre-worker-exit hook validates this file and blocks the worker if anything is missing. Activate inside any worker subagent.
Per-feature 20-question interrogation that turns a draft feature spec into a fully-specified one. Three rounds — 10 task questions, 5 follow-ups, 5 "definition of done" questions. Each question has 4 options with one ★ recommended based on context. Activate during /mission-tasks.
| name | worker-mcp-usage |
| description | When and how workers use MCP tools during /mission-run. Prefer MCP for live service introspection (Supabase schema, GitHub, etc.) when connections/mcp-registry.md lists the service. Activate in worker subagents during execution. |
Workers implement features in /mission-run. External services were registered
during /mission-connect. The orchestrator never asks the user questions
during run — resolve ambiguity from clarified specs and this skill.
missions/<id>/connections/mcp-registry.md — which MCP servers exist and tool prefixesmissions/<id>/tech-decisions.md — SDK names, env var names, conventions| Situation | Use MCP | Use SDK/code in repo |
|---|---|---|
| Inspect live schema, tables, columns, RLS policies | Yes (e.g. Supabase MCP) | No for one-off checks |
| Apply or verify migrations against real project | Yes when MCP supports it | Migrations files in repo + tests |
| Query seed/fixture data to validate a feature | Yes | — |
| Implement product API routes, components, business logic | No | Yes |
| Unit/integration tests in the test runner | No | Yes |
| Stripe/Resend/etc. with no MCP in registry | No | SDK + env from .env |
Rule: If mcp-registry.md lists the service with Worker use: yes, and your
feature touches that service's live state, call the MCP tools before or after
code changes to verify schema, policies, or remote configuration. Record which
tools you used in the handoff Commands run / Decisions made sections.
Registry rows include Tool prefix (e.g. mcp__supabase__*). Use tools matching
that prefix. If a tool call fails with "not found", set handoff Status to
BLOCKED or PARTIAL with:
BLOCKER: MCP tool unavailable for <service>
TRIED: <tool names attempted>
NEEDED: Orchestrator should re-run /mission-connect or verify claude mcp list
SUGGESTED FOLLOWUP: <one paragraph to register/fix MCP for this service>
Do not ask the user questions. The orchestrator handles recovery without stopping the run.
.env via the project's standard mechanism — never log valuesPriority order when the spec is unclear:
clarifications/F<NNN>-clarification.mdtech-decisions.md conventionsBLOCKED with SUGGESTED FOLLOWUP (orchestrator creates follow-up feature)Optional in handoff when you applied a default:
AUTONOMOUS_DECISION: <what you chose and why>
Validators do not see your chat. Link MCP verification in handoff:
Decisions made: e.g. "Verified users table and RLS via Supabase MCP list_tables"Commands run: note significant MCP tool invocations (tool name only, no secrets)