What problem does it solve? Reviewing skill plugins with a single LLM misses semantic issues like logic bugs, citation drift, and undocumented prerequisites that a deterministic linter cannot catch, and single-model reviews inherit that model's blind spots. ## Core Features & Use Cases - Parallel multi-model review: Spawns Claude Agent personas, Codex GPT-5.5, and Cursor Composer 2.5 against the same per-plugin prompt, then aggregates findings with per-finding agreement strength (4/4 down to 1/4). - Solo-mode fallback: When external CLIs are unavailable, falls back to three Claude personas (Security/Infrastructure, Usability/Documentation, Cross-skill Consistency) with an explicit warning about weaker same-model evidence. - Structured severity classification: Tags every finding as (a) deployment-grounded, (b) usability-grounded, or (c) preference, mapping to P0/P2 actions. - Use Case: Before merging a new plugin to the fgcz-skills marketplace, run this skill to get cross-model consensus on utility verdicts and P0 findings, then append the result to the review history index. ## Quick Start Run a multi-LLM review of the bfabric-lims plugin and aggregate the findings into a consensus report.
