orchestrator
| Type | Skill |
| Plugin | awl-general · v0.0.51 |
| Invoke | /awl-general:orchestrator |
| Source | plugins/awl-general/skills/orchestrator/SKILL.md |
When Claude uses it
Section titled “When Claude uses it”Orchestrate implementation using parallel subagents or agent teams. Use when implementing plans, features, or multi-file changes.
Definition
Section titled “Definition”Core Principle: This context window is for orchestration ONLY. Offload ALL work to subagents or team members.
Mode Selection
Section titled “Mode Selection”Choose orchestration mode based on task complexity:
| Condition | Mode |
|---|---|
| Independent tasks, no cross-talk needed | Subagents (parallel Task calls) |
| Workers need to coordinate or share findings | Agent Team (TeamCreate workflow) |
| < 3 tasks total | Subagents (team overhead not worth it) |
| Cross-layer work (frontend + backend + tests) | Agent Team |
Default to subagents unless coordination is clearly needed.
Mode A: Subagent Orchestration
Section titled “Mode A: Subagent Orchestration”Execute stages in order. Parallelize within stages. Use the Task tool to spawn agents — pick the subagent_type that best fits each job from the available agent types (see Task tool description for the full list).
1. Implement
Section titled “1. Implement”Spawn parallel Task agents per file/component. Write detailed prompts with:
- Exact file paths to create/modify
- Full requirements and acceptance criteria
- Any relevant context (existing patterns, types, interfaces)
Spawn all independent implementation tasks in a single message for maximum parallelism. Wait for all to complete before next stage.
2. Build
Section titled “2. Build”Single agent to verify the build passes. Fix any compilation/type errors before proceeding.
3. Test
Section titled “3. Test”Spawn parallel agents per test scope. Each agent writes tests for one component/module.
4. Review
Section titled “4. Review”Single agent reviews all changes. Address any critical findings.
5. Verify
Section titled “5. Verify”Single agent verifies plan completeness — checks each plan item against the codebase.
6. Simplify (Optional)
Section titled “6. Simplify (Optional)”Only if code is complex or reviewer flagged issues.
Mode B: Agent Team Orchestration
Section titled “Mode B: Agent Team Orchestration”Use when tasks require inter-agent coordination, cross-layer work, or when the user explicitly requests a team/swarm.
- TeamCreate — initialize team with name and description
- TaskCreate × N — define ALL work units upfront
- Write detailed descriptions — teammates read these as their spec
- Set dependencies with
addBlockedBywhere needed
- Task × N — spawn teammates with
team_nameparameter- Each teammate gets a full context window
- Use
model: "sonnet"for teammates to optimize cost (lead stays on opus)
Teammate Prompt Template
Section titled “Teammate Prompt Template”Each teammate spawned via Task should receive:
You are a teammate on team "{team_name}". Your role: {role}.
Check TaskList for available tasks. Claim unblocked tasks by setting yourself as owner.When done, mark completed and check TaskList for next work.Use SendMessage to report findings or ask for help.Execution Loop
Section titled “Execution Loop”The lead (you) monitors progress:
- Teammates auto-claim tasks from shared TaskList
- Messages from teammates arrive automatically
- Assign tasks with TaskUpdate if teammates need direction
- Use SendMessage for direct coordination
Teardown
Section titled “Teardown”SendMessage({ type: "shutdown_request" })to each teammate- Wait for all
shutdown_responseapprovals TeamDelete()
- Never read/write files directly — delegate to agents/teammates
- No git operations — never commit, push, or perform any git actions
- Maximize parallelism — spawn multiple agents when tasks are independent
- Detailed prompts — include all requirements in the prompt for clarity
- Report progress — summarize each stage completion to the user
- Handle failures — if agent fails, spawn replacement with error context
Example: Subagent Mode
Section titled “Example: Subagent Mode”User: "Implement the login feature from the plan"
Orchestrator:1. Spawn 3 implementation agents in parallel (single message): - LoginViewModel - LoginScreen - AuthRepository2. Wait for completion3. Spawn agent to verify build4. Fix any issues5. Spawn 2 test agents in parallel6. Spawn review agent7. Spawn agent to verify plan completeness8. Report final statusExample: Team Mode
Section titled “Example: Team Mode”User: "Use a team to build the checkout feature"
Orchestrator:1. TeamCreate("checkout-feature")2. TaskCreate: backend API, frontend UI, integration tests, e2e tests3. Set dependencies: tests blocked by implementation4. Spawn 3 teammates: backend-dev, frontend-dev, qa-engineer5. Monitor progress via automatic messages6. When all tasks complete: shutdown teammates, TeamDelete7. Report final status
