メニュー
Coordinate multiple Claude Code sessions working on the same project via shared task list. Creates task DAG, assigns work to parallel agents, monitors progress. Use for large implementations that benefit from true parallelism.
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
SOC 職業分類に基づく
| name | swarm-coordinator |
| description | Coordinate multiple Claude Code sessions working on the same project via shared task list. Creates task DAG, assigns work to parallel agents, monitors progress. Use for large implementations that benefit from true parallelism. |
| context | conversation |
| skills | ["writing-plans","task-builder","requesting-code-review"] |
Orchestrate multiple Claude Code sessions or agents working on the same project through a shared task list.
The swarm coordinator uses CLAUDE_CODE_TASK_LIST_ID to enable multiple Claude Code instances to share progress state. Each instance can:
TaskListParallel: yes tasksTASK_LIST_ID="swarm-$(date +%Y%m%d-%H%M%S)"
echo "export CLAUDE_CODE_TASK_LIST_ID=$TASK_LIST_ID"
Save this ID - all workers need it.
If IMPLEMENTATION_PLAN.md exists:
TaskCreate, recording plan ID → system ID mappingTaskUpdate using mapped system IDsOutput the task graph for user review.
Guide user to open additional terminal windows:
# Terminal 2 (worker)
export CLAUDE_CODE_TASK_LIST_ID=[same-id]
cd [project]
claude
# Terminal 3 (worker)
export CLAUDE_CODE_TASK_LIST_ID=[same-id]
cd [project]
claude
Each worker follows this loop:
1. TaskList to see available work
2. Find unblocked pending task with no owner
3. TaskUpdate to claim (set owner to session ID)
4. TaskUpdate status to in_progress
5. Implement the task
6. Run tests
7. TaskUpdate status to completed
8. Repeat
The coordinator (this session) monitors progress:
- TaskList every few minutes to see global state
- Identify stuck tasks (in_progress too long)
- Handle blockers and conflicts
- Run final integration tests when all tasks complete
To prevent race conditions:
TaskUpdate with owner: "session-<id>" for a pending taskTaskGet to verify ownership before proceedingin_progress and completedIf two workers modify the same file:
When all tasks are completed:
/requesting-code-reviewCoordinator: "Starting swarm for auth-system implementation"
TaskList shows:
#1 [pending] Create user model - ready
#2 [pending] Create auth middleware - ready
#3 [pending] Create login endpoint - blocked by #1, #2
#4 [pending] Create register endpoint - blocked by #1
#5 [pending] Add password hashing - blocked by #1
Coordinator assigns:
- Worker 1: Task #1 (user model)
- Worker 2: Task #2 (auth middleware)
Workers complete, TaskList now shows:
#1 [completed] - by worker-1
#2 [completed] - by worker-2
#3 [pending] - now unblocked!
#4 [pending] - now unblocked!
#5 [pending] - now unblocked!
Coordinator assigns next batch...
CLAUDE_CODE_TASK_LIST_IDUse when completing tasks, implementing major features, or before merging to verify work meets requirements - runs review in a forked code-reviewer agent
Activate autonomous loop mode for persistent development sessions. **CRITICAL:** Always activate this AFTER completing spec and implementation plan, BEFORE beginning implementation. This is what makes autonomous builds actually autonomous—the Stop hook enforces completion criteria.
Use when creating or developing, before writing code or implementation plans - refines rough ideas into fully-formed designs through collaborative questioning, alternative exploration, and incremental validation. Don't use during clear 'mechanical' processes
Apply systematic root cause analysis and debugging methodologies to diagnose and fix bugs, test failures, and unexpected behavior. Use when encountering production issues, investigating test failures, diagnosing performance problems, tracing error sources through call stacks, analyzing logs and stack traces, reproducing inconsistent bugs, debugging race conditions, investigating memory leaks, or applying scientific method to problem-solving before proposing fixes.
Activate orchestrator mode. You coordinate specialists, you don't implement. MAXIMIZE parallelism with task-builder. MAXIMIZE quality with code-reviewer, codex, gemini. Speed + Quality through delegation, not through doing work yourself.
Execute ONE task in parallel with other task-builders. ALWAYS SPAWN MULTIPLE for independent tasks. ORCHESTRATOR: You coordinate, you don't implement. Spawn N task-builders for N unblocked tasks. Maximum parallelism = maximum speed = maximum quality (each task gets focused attention).