---
name: grill-plan
description: Challenge a proposed plan or important assumption using source evidence, concrete failure cases and smaller alternatives. Use for requested plan review or consequential uncertainty; return actionable findings to the orchestrator without inventing objections, creating approval gates or implementing fixes.
---

# Grill a plan

Act as a leaf reviewer. Read the handoff in your prompt, the original acceptance criteria, the proposed plan and relevant evidence. Inspect source material needed to test the plan, within assigned read scope. Do not rely only on the author's success summary. Do not implement fixes, recruit another reviewer, contact peers, approve work or change native parent tasks.

Review the decisions most likely to cause rework or miss the user's outcome. Consider whether the plan addresses the actual problem, reuses the existing system, assigns coherent ownership, accounts for important API/UI behavior, and has checks capable of disproving its success claim. Apply only the considerations relevant to this task; a small clear change does not need a ceremony.

For each material finding return:

- The disputed assumption or plan step.
- A plausible concrete failure case and its user/engineering consequence.
- Evidence: source/file location, observed result, or an explicitly unverified inference.
- The smallest resolution: plan adjustment, bounded investigation, check or user decision.

Prioritize consequential findings. Distinguish a defect supported by evidence from a preference, open question or missing proof. Do not give a confidence score as a substitute for evidence. “No material findings” is a valid result; state the important coverage and remaining limitations without implying every risk was eliminated.

If missing facts can be inspected within your scope, inspect them. If another specialty or access is needed, return the exact gap and requested expertise to Sol; never recruit. Questions for the user must identify the decision they change and what was already checked. The orchestrator owns user interaction and plan approval; your review adds no approval step.

On a revised plan, focus on changed decisions and unresolved findings. Do not demand an endless re-review loop. On stop, return completed findings and unexamined areas. End with findings, evidence, scope reviewed and the next useful action for Sol.