grilllisted

# /grill — Adversarial Self-Check **Trigger:** After completing any unit of work. Before claiming done. ## Procedure ### Step 1: Ask four questions Answer each against the most recent work product. Be concrete — name files, lines, behaviors. 1. **"Where does this break?"** — Edge cases, empty inputs, concurrency, platform differences, missing data. 2. **"What have I not checked that I should have?"** — Files not read, paths not traced, integration points not tested. 3. **"What's the most likely thing I got wrong?"** — The answer is never "nothing." Overconfidence is the signal to look harder. 4. **"What assumption am I making that I haven't verified?"** — Constants, file existence, enum completeness, spec accuracy, API behavior. ### Step 2: Check every checkable answer For each answer that points to a specific file, value, behavior, or path: go verify it now. Read the file. Run the grep. Trace the call site. Do not reason from memory. Skip only answers that are genuinely uncheckable (e.g., "a user might dislike this UX" — that requires user testing, not a file read). ### Step 3: Resolve or flag - If a check reveals a problem: fix it before proceeding. - If a check confirms correctness: note it briefly and move on. - If all four questions yield "nothing identified" after honest effort: done. ### Step 4 (optional): Prove it works When the work is a behavior change: diff the before and after. Show the concrete difference in output, state, or user experience. "It work