contrarylisted

# Contrary Mode Adversarial critic, delegated. Do NOT write code. Do NOT edit files. Do NOT run tests. Do NOT critique inline yourself — your job is to package the target and hand it to a fresh subagent. Output goes to chat only. ## Target from user "$ARGUMENTS" If empty, ask the user what they want challenged. Stop until they answer. ## Process ### 1. Distill a self-contained brief The subagent has zero memory of this session. Whatever you send must stand alone. - If the target is a file path, PR URL, branch, or doc — include the exact path/URL so the subagent can read it. - If it's a free-text claim or plan — restate it in full. - If it references prior conversation ("the approach we just decided on", "the plan above", "what we just built") — extract the specific claim/plan/decision, the relevant surrounding context, and any files/PRs/URLs the subagent will need. Don't say "the thing we discussed" — spell it out. Attacking a strawman is worse than not attacking at all. If you're unsure what the user means, ask before dispatching. ### 2. Dispatch the critic — default to the other LLM Default: run the critique via the `other-llm` skill, in review mode — a genuinely different model family is a more independent adversarial read than one model critiquing itself. Follow the `other-llm` skill's process, using the critique instructions below (prefixed with its own "READ-ONLY REVIEW" line per that skill's Step 2) as the prompt. Fallback: if the other CLI is unavailable or i