compose-skillslisted
Install: claude install-skill sumitake/agent-collab
# Compose skills - bounded multi-skill planning
Choose the smallest useful set of collaboration skills or lenses for a task. This skill is advisory: it plans how to combine skills, but it does not execute the selected skills, route agents, or override governance.
## Operating contract
- Prefer 1 primary lens, 1 supporting lens, and at most 1 verifier unless the task proves it needs more.
- Load metadata first. Load full skill bodies only when that skill is about to be used.
- Treat recipe files as planning hints, not routing policy. Agent routing stays with the workspace routing policy; model and effort selection stay with the model-effort policy.
- Keep token cost explicit. More agents, longer context, and chained reviews must have a reason tied to the task's risk or uncertainty.
- If the task involves untrusted external material, include `untrusted-audit` before any adoption, execution, hook installation, or workflow import.
## Workflow
1. Restate the task outcome in one sentence.
2. Identify the minimum lens set:
- **Primary lens:** the skill that does the core work.
- **Supporting lens:** a skill that supplies missing perspective, such as `brainstorm`, `long-context`, `delegate`, or `code-review`.
- **Verifier lens:** a skill that checks the result, such as `qa-verify`, `logic-check`, `red-team`, or `second-opinion`.
3. If the workspace contains `workspace/config/skill-composition-recipes.yaml`, run or consult its validator before relying on a named recipe: