Review the latest version of the plan, preserving the original goals, scope, and reasonable design without redoing requirements interviews. First check project assets and documentation without repeating already answered questions; if a previous plan is missing, request it.
1. Grill: Implicit Decisions (Required)
Identify hidden decisions the plan has implicitly made for the user or high-consequence behaviors it has omitted. Follow the user's workflow to examine defaults, error/cancel/retry paths, state recovery, cross-module contracts, compatibility migrations, and permission boundaries. Challenge assumptions with concrete counterexamples rather than mechanically running a checklist.
Only ask about real tradeoffs that would significantly affect experience, architecture, data, security, or cost; resolve factual or low-cost reversible details on your own. For each question, state the assumption/gap, consequences, and a suggested path. Multiple rounds are allowed; ask many questions per round, grouped by topic, starting with prerequisite decisions. Wait for user responses before finalizing; silence does not imply consent.
2. Oracle: Objective Validation (Optional)
Enable only when credible, cost-effective objective criteria exist; the user may skip this. For each critical requirement, specify the criterion source, inspection method, and pass condition; prioritize specifications, reference implementations, user-confirmed examples, and invariants. For optimization tasks, list baselines, metrics, and constraints that must not regress. Do not treat model self-evaluation or agent metrics as ground truth; parts that cannot be automatically verified retain manual acceptance.
3. Prototype: Subjective Comparison (Optional)
Enable only when appearance or interaction quality influences decisions and text alone is insufficient; the user may skip this. Use references or a small number of minimally different prototypes for comparison, without continuing to probe abstract preferences. Only create isolated prototypes if the current mode permits and the user has authorized it; otherwise, write "Prototype comparison → User selection → Related implementation" into the second edition, leaving corresponding open items. Do not claim validation without showing or comparing.
Convergence
Converge after important decisions are confirmed, explicitly delegated, or deferred; also converge when the user requests to end. Output a complete, independently usable second-edition plan that integrates decisions, implementation steps, and applicable validation/prototype gates, along with a brief summary of changes and remaining assumptions; for deferred items, note when the decision is needed and what it blocks. Do not deliver only a questionnaire or patch; do not claim to have exhausted all risks. After output, stop; do not self-implement or restart the review.