| name | grill-plan-review |
| description | Review a long or cognitively heavy plan through a concise, one-decision-at-a-time interview, recommend each choice with rationale, and turn the user's answers into a persisted accepted plan and execution authorization when requested. Use when the user cannot or does not want to read a plan document, asks for a grill-me-style plan review, or wants interactive confirmation before implementation. |
Grill Plan Review
Replace document reading with a decision interview while preserving a precise
approval trail.
- Read the complete proposed plan, task status, and material evidence first.
Resolve discoverable facts yourself instead of asking the user.
- Extract only decisions that materially change scope, safety, cost,
reversibility, sequencing, validation, or stopping conditions.
- Use
request_user_input for one focused decision at a time with 2-3
concrete options. Put the recommended option first, label it
(Recommended), and explain its rationale and tradeoff in the description.
- After every answer, state the locked implication in one or two sentences,
then ask the next unresolved decision. Keep an explicit decision ledger.
- Before the final confirmation, summarize the accumulated safeguards
concisely. Ask whether to save only or save and execute unless the user has
already made the intended post-review action unambiguous.
- Treat the answers as review/confirmation only when the user explicitly said
they should count that way. Preserve the exact answers as immutable task
input.
- If the final selection authorizes execution:
- persist the full accepted plan in the canonical task status file before
implementation;
- include timestamp, accepted signal, supersedes value, revision summary,
full plan, acceptance criteria, and hard boundaries;
- create or update the requested goal;
- execute without another redundant approval prompt.
- Stop for clarification instead of inferring approval when the interview
exposes a new action outside the reviewed scope.
Keep the interaction easy to scan: lead with the current decision, avoid
repeating the long source document, and make costs and failure behavior
concrete.