named-persona-adversarial-review

Pick personas per the routing table (default: Torvalds, Thompson, Jobs), ground each in their documented principles from references/persona_principles.md before reading code, then have each produce findings mapped to a principle with a confidence level. Synthesize duplicates, promote concurrent findings, and post the report as a PR comment.

Abstract adversarial review uses generic roles like 'saboteur', which produces surface-level findings. Named-persona review anchors each lens in a real, sourced philosophy (e.g., Thompson's Reflections on Trusting Trust), so findings cite a defensible principle rather than an imagined opinion.

It enforces an attribution discipline: cite the documented principle with a confidence level (high/moderate/low) instead of inventing verbatim quotes. If a persona's lens cannot be grounded in references/persona_principles.md or a verifiable source, that persona is dropped from the round.

For security and supply chain, use Thompson plus Torvalds plus Cagan. For performance, use Carmack plus Torvalds plus Jobs. The routing table in SKILL.md maps problem types to persona mixes, and a product persona is mandatory in every round.

Skip it for cosmetic or low-impact PRs (use the simpler adversarial-reviewer instead), for throwaway prototype code, and when you have no web access and the persona is not covered in references/persona_principles.md, since ungrounded personas must not be used.

One round minimum for any PR. If BLOCKER or CRITICAL findings appear, fix them and run one re-review round. Two consecutive clean rounds end the process, and a single clean round suffices for low-impact PRs.