Grill Me

A plan that survives only its author's scrutiny is untested. grill-me puts the plan in front of a different model whose job is to break it, not bless it.

When to use

• An architecture or design proposal is internally consistent but unchallenged.
• You have only one model's context on a high-stakes decision.
• A plan keeps "working" in tests but you can't name its failure modes.

Workflow

Step 1 — Planner emits the design

Model A (e.g. Claude) produces: the design, the explicit assumptions it rests on, and the acceptance criteria. No defense yet — just the artifact.

Step 2 — Challenger attacks

Model B (e.g. Codex) receives the design + assumptions and attacks:

• Assumptions that don't hold under load / edge input / partial failure.
• Failure modes with no mitigation.
• Cheaper alternatives the planner dismissed without evidence. The challenger may not rewrite the design — only test it.

Step 3 — Planner responds or concedes

Model A addresses each challenge: fix, accept-and-mitigate, or concede the flaw. Conceded flaws become explicit risks. Repeat ≤ 3 rounds; stop early if the challenger repeats itself.

Step 4 — Verdict

Final design + a list of survived vs. refuted challenges + residual risks the builder must watch. The verdict is the deliverable, not a "looks good".

Anti-patterns (never do these)

• Letting the challenger become a yes-man — rotate models if it concedes too fast.
• Running rounds with no convergence — cap at 3 and summarize the standoff.
• Challenger rewriting the design — that's a second planner, not a grill.
• Grilling a trivial decision — reserve for high-stakes, hard-to-reverse calls.

Output contract

Return: the final (or standoff) design, a two-column table of challenges (survived / refuted), and a bulleted residual-risk list. If the two models cannot converge, say so and hand both views to the human. See references/challenge-prompts.md for the challenger's angle set.