multi-substrate-reviewlisted

# Multi-Substrate Review / Fanout The most-repeated high-value move this operator makes: get ≥1 NON-Claude substrate to independently review/attempt something, then verify and synthesize. A single frozen model can't surprise itself; heterogeneous substrates catch each other's blind spots. This skill is that pattern, packaged. ## When to use - Reviewing your OWN work (code, design, contract) before it lands — you are the worst reviewer of your own output. - Any consequential decision/claim where one model's confidence is a risk. - Generating N independent attempts to synthesize the best. ## The substrates (verified available 2026-06-05) | substrate | invoke (read-only / non-interactive) | strength | accuracy on code-AUDIT | |---|---|---|---| | **codex** (gpt-5.5) | `codex exec -s read-only --skip-git-repo-check "PROMPT"` | reasoning, audit | HIGH (found 3 real bugs Claude missed) | | **gemini** | `gemini -p "PROMPT" --approval-mode plan` | audit, breadth | HIGH (found the enforcement gap) | | **qwen3-coder:30b** (local) | `hivemind/.local/ollama/bin/ollama run qwen3-coder:30b "PROMPT"` or `:11434/v1` | gen, tool-use, bulk, private | LOW (4/5 false positives — see [[local-llm-agent]]) | Routing rule (measured): **audits/reviews → codex &/or gemini; generation/draft/ bulk/private → local qwen3-coder.** Always Claude-verifies every substrate's output — substrates propose, Claude disposes. ## Workflow 1. **Frame one sharp prompt** — concrete target, "find what I missed",