Adversarial Review

Sends a deeply structured adversarial prompt to a heavyweight reviewer (the codex frontier model or Opus 4.7) to catch real bugs before they hit production. The reviewer reads actual file contents, not summaries, and returns a trinary verdict with file:line citations and a prioritized fix list.

Distinct from RedTeam (which adversarially challenges plans and ideas with parallel expert attackers) and grilling (which interviews the user about designs). This skill reviews working or near-complete code against its own specification.

Workflow

Step 1 — Gather inputs

Use AskUserQuestion with these four questions in a single prompt:

1. Target directory — What directory should the reviewer work in? (e.g., ~/Projects/arbol/v3)
2. Plan or spec file — Path to the spec, plan, or design doc to review against. If none, the reviewer will assess internal consistency instead.
3. Reviewer — Which model? Options:
   • the codex frontier model (recommended — deepest code reasoning, runs locally)
   • opus 4.7 (alternative — best for architecture-level concerns)
4. Review areas — Provide 4–10 numbered areas to focus on (e.g., "1. Cron scheduling, 2. File path assumptions, 3. Error handling"). If unspecified, use the 8 default areas from the prompt template.