adversarial-reviewlisted
Install: claude install-skill ucsandman/claude-skills
# Adversarial Review
Review a change-set with a two-stage multi-agent pipeline: **find** (parallel
read-only reviewers, each owning one failure dimension) then **verify** (an
independent skeptic per finding, prompted to refute it). Only findings that
survive verification are reported. This kills the two classic failure modes of
LLM review: plausible-but-wrong findings, and one reviewer trying to hold every
concern at once.
This skill authorizes use of the Workflow tool for the orchestration.
## 1. Establish scope (cheap, inline)
- Default: the working tree — `git diff HEAD --stat` plus `git status --short`
for untracked files. If the tree is clean, use commits ahead of the default
branch (`git log origin/main..HEAD`); if none, the last commit.
- If args name a base ref, PR number (`gh pr diff <n> --stat`), or paths, scope
to that instead.
- Build a SCOPE block: repo path, the changed-file list grouped by
feature/area if discernible (read the recent commit messages), and one line
of context per group ("X was just implemented by engineer A who did not know
about Y"). If several independent changes landed in sequence, SAY SO in the
scope — it unlocks the seams dimension.
- Note anything reviewers must NOT do (modify files, run test suites that
would race a separate verification run).
## 2. Pick dimensions
Default four; drop or add based on what the diff actually contains:
| Dimension | Hunts for |
|---|---|
| correctness | real bugs: off-by-one, wrong clam