What problem does it solve? Independent review loses value when the reviewer shares the author's session context and prior reasoning. This Skill obtains genuinely cold reads by sending a mandate and documents to a fresh model instance with no memory of the current session, capturing the reply verbatim with a sidecar recording exactly what ran. ## Core Features & Use Cases - Multi-model roster: One request format reaches Claude subagents, GPT via codex exec, Gemini via an MCP tool, and DeepSeek or Qwen via OpenRouter, with per-row models, efforts, and access profiles resolved from a roster file. - Verbatim evidence capture: Every call writes raw.md, a sha256 hash, and a sidecar.json with status, effective model, budget, and parity lines, so review output is auditable and never edited. - Pre-send validation and guards: Requests pass authorization, profile, floor, budget, and lane-availability checks before anything is sent; refusals and failures are recorded as first-class results. - Use Case: During an adversarial review of a build doc, a caller skill summons a fleet of readers sharing one run id — a fresh Claude subagent plus GPT and Gemini — and receives each cold read verbatim with sidecars proving isolation. ## Quick Start Ask the assistant to run /readers claude-opus path/to/design-doc.md "Review this design doc for correctness gaps" to get one cold read captured verbatim with its sidecar.
