cross-model-review

One-shot cross-model verdict — the same open question goes to Claude and to Codex, each answering cold, then one peer-critique round, then a double-blind claim-level verdict on which answer holds up. Triggers on "ai-duel", "get a second opinion", "ask another model", "what would a different model say", "settle this for me", "which approach is better" — or any design choice with no textbook answer where both paths are defensible. NOT for iterating to consensus (use xreview), fact and literature lookup, poking holes in a piece of writing (use ai-review), or a coding task where two models would edit the same file (isolate them in two git worktrees instead).

Anti-flattery manuscript review — a piece of writing goes through anonymisation plus a multi-lens review: four lenses in parallel under an absolute ban on praise, a rebuttal round that filters out criticism invented to look thorough, then an editor pass that produces a revised draft. Triggers on "ai-review", "review this draft", "poke holes in this", "what's wrong with this piece", "don't just tell me it's good", "be honest about this", "tear this apart". NOT for plans and design proposals (use xreview), choosing between two options (use ai-duel), code review, or fact checking.

Cross-model review loop — hand a plan, proposal or script to a different model family (Codex by default; Claude or Grok as alternatives) as the reviewer and trade rounds inside one reviewer session, fixing what should be fixed and pushing back where it is wrong, until both sides genuinely reach consensus; only then is the XREVIEW-PASS marker stamped. Use when the user says 'cross-review this', 'have another model review this', 'get Codex to check this', 'get Grok to review this', 'review it until we agree', 'xreview', or before implementing a non-trivial plan, before an expensive computation run, before writing AI output into any trusted store (reference library, notes, knowledge base), or before a major method or architecture choice. NOT for: one-line changes; plain factual lookups; a one-shot 'who is right' ruling (use ai-duel).