SKILL.md
Parallel Launch
Decompose the given task into independent agent groups and execute them in broad parallel.
Process
- Analyze the task and identify independent concerns that can run concurrently.
- Each concern must be self-contained: no shared mutable state, no ordering dependency. - If concerns have dependencies, sequence the dependent batch after the independent batch completes. - Consult [delegation scenarios](./references/delegation-scenarios.md) for parallelism decisions.
- Design agent groups — for each independent concern:
- Assign a clear, scoped objective (one concern per agent). - Select the appropriate agent type (Explore, Plan, general-purpose, or domain specialist). - Define expected output format so results can be composed.
- Launch all independent agents in a single tool call — never sequentially when parallel is possible.
- Compose results once all agents complete:
- Merge non-conflicting outputs directly. - For conflicting or overlapping results, reconcile and present trade-offs to the user. - If any agent failed or returned incomplete results, report the gap and propose a targeted follow-up.
- Review composed output — dispatch a review agent to verify:
- Completeness: All original concerns addressed, no gaps. - Consistency: No contradictions between agent outputs. - Accuracy: Claims are substantiated, sources checked, no hallucinated findings. - Scope: Nothing extra built beyond what was asked. - For implementation work, additionally verify spec compliance and code quality.
- Report to user only after review passes.
Constraints
- Agents per batch: match the number of truly independent concerns (avoid artificial splitting).
- Each agent prompt must include full context — agents do not share memory.
- Do not launch agents for trivially sequential work (single file, single concern).
- If the task has fewer than 2 independent concerns, execute directly instead of launching agents.
Red Flags
- Never skip review. Composed output must always pass through a review agent before reporting.
- Never accept unverified composed output. If agents return conflicting results, the review agent must flag them — not silently pick one.
- Never report to user before review passes. The review step is mandatory, not advisory.