panellisted

# panel `prove` checks numbers by recomputing each from the source of truth. `grill` sends a single attacker after a claim. `panel` runs several expert views, each judging the same decision independently. It then states what they agree on, where they truly conflict, and one recommendation. Use it when no single perspective sees the full picture. ## When to use it - A design, architecture, or risky call that needs several disciplines. - A decision where a conflict between views (security vs. speed, user need vs. maintainability) is the real finding, not noise to smooth over. - Skip it when a single correctness check or an adversarial test covers the question. ## Required inputs - The decision, design, or artifact to assess. - Optional: preferred views; otherwise this prompt selects 3 to 5 that fit. Common choices: correctness, security, performance, maintainability, user, skeptic. - Any spec, constraints, or context a view needs to judge fairly. ## Steps 1. Name the views before assessing. Commit to 3 to 5 that match the decision. 2. Run the views. Each judges from its own perspective only: what it accepts, what concerns it, and what would change its verdict. On a host that can run subagents, run the views in parallel so no view can see another. Otherwise run them in sequence and write each view's full take before reading the previous ones back; that states the independence intent honestly, but a later view may still anchor on an earlier one, so the report