| name | light-plan-and-work |
| description | Create a 3–7 step plan for a bounded task, execute immediately, verify, and hand it back. Use when the user explicitly requests a lightweight plan-and-work flow; do not use for open-ended brainstorming, discovery, high-risk migrations, multi-system architecture, or releases. |
| disable-model-invocation | true |
Light Plan and Work
Planning and doing are one continuous workflow.
Route before planning
Read the request, relevant project instructions, current files, and repository state. Choose one route:
- Direct execution: one or two obvious actions. State a one-line brief and do them; do not manufacture a plan.
- Light plan and work: the outcome is concrete, the scope is bounded, and 3–7 steps can complete and verify it.
- Specialist Skill: a narrower installed Skill owns the artifact or domain. Use it, with this Skill only as the execution wrapper when useful.
- Discovery or brainstorming: the user is still choosing the problem, audience, concept, story, or direction. Use a discovery Skill before planning.
- Heavy workflow: use the project's full specification or plan/work system when consequence, ambiguity, or coordination cost is high.
Heavy conditions include destructive or sensitive operations, migrations, public API changes, cross-system architecture, releases, multiple owners, unresolved acceptance criteria, and explicit requests for a full specification or Compound Engineering. Read routing and verification when the route is unclear.
Execute
- Lock a compact execution brief: Goal, Deliverable, Boundary, and Acceptance.
- Resolve reversible choices yourself. Ask one blocking question only when the answer changes the outcome or requires new authority.
- Use the host plan tracker for 3–7 observable steps, with at most one step in progress. Each step must produce or verify something.
- Start after the plan is visible. Read before editing, preserve unrelated work, follow project instructions, and keep the plan aligned with evidence.
- If a heavy condition appears, preserve completed safe work, pause the affected mutation, explain the evidence, and switch to the heavier workflow.
- Verify in proportion to risk. Distinguish checks that directly cover the task from repository-wide or environment-wide gates.
- When a broad gate fails, classify the failure before acting. If read-only evidence shows it comes from pre-existing, unrelated work, preserve that work, run the strongest safe task-scoped checks, and record the blocked gate without claiming the repository is fully green. If the failure overlaps task files, causality is uncertain, or the gate is required for a release, treat it as a task blocker and fix, pause, or escalate.
- Hand back the result, artifacts, evidence, blocked gates, and residual risk.
Never revert, stash, edit, or stage unrelated work to make a broad verification command pass. Read routing and verification for the evidence and handoff contract.
Do not create a durable plan file by default. Create one only for an explicit request, future resumption, external handoff, multiple owners, or a lasting design decision, and follow the project's existing plan location.
Maintain trigger boundaries with evals/trigger_cases.json and output behavior with the baseline comparisons in evals/output/cases.jsonl.