I stopped trusting Claude's code reviews, so I built a skill that puts my code on trial

A developer built Tribunal, an adversarial Claude skill that reviews code diffs by pitting biased agents against each other to produce more honest feedback. The system uses per-file 'haters' that attack the diff on technical merits and a cross-module bug hunter, with a judge that validates each accusation against actual code intent. The skill is open-source, portable across Claude Code and Claude Cowork, and supports multiple programming languages.

Every time I asked Claude to review my branch, I got one of two answers: a cheerful "Looks good 👍" or a vague list where I couldn't tell a real bug from a matter of taste. The model wants to please you. That's exactly the problem. So I built Tribunal — a Claude skill that reviews your diff adversarially , in stages, where the honest signal comes from agents fighting each other instead of one polite model. A single model told to "be critical" still hedges — it's trained to be agreeable. So instead of one balanced reviewer, Tribunal runs one-sided roles that collide : One agent per file, deliberately biased. It tears the diff apart as if a clueless amateur wrote it — focused only on what changed. But strictly on the merits: correctness, races, leaks, edge cases, security. No style nitpicks. Per-file haters are blind to cross-module bugs. A separate agent hunts exactly those: a changed function signature whose caller still calls the old way, a return shape a consumer no longer matches, invariants out of sync across files. For each accusation, the judge digs into the actual code and decides honestly: was this deliberate and justified , or genuinely weak ? It's allowed to use docs and comments as evidence of intent — the opposite of the hater, who ignores them as excuses. Keeps only the spots the judge couldn't defend — or conceded are weak even while defending the choice. Everything else drops to a full transcript. The balance doesn't live inside any single agent — it comes from the clash between them. A hater that can only attack, meeting a judge that only looks for justification, produces a sharper, more honest signal than one model trying to be "balanced" on its own. And the hater is allowed to return nothing. On a clean diff it's not forced to invent problems — empty is a valid, honest result. A ranked report written to docs/reviews/ , plus a short chat summary: what to actually fix, by severity critical → major → minor , with a concrete fix for each. It's portable — pure Claude sub-agents the Agent tool , no external runtime, no dependencies. Works in Claude Code and Claude Cowork , in any language Python, JS/TS, Go, Rust, Java… one config line to add yours . It's MIT and free: https://github.com/hekman316/claude-skill-tribunal https://github.com/hekman316/claude-skill-tribunal Install is one paste — ask Claude to fetch the SKILL.md from the repo and drop it in ~/.claude/skills/ . Then in any repo just say /tribunal . I'm genuinely curious what people think of the adversarial-roles approach. Does forcing the model into one-sided roles actually beat just asking it to be harsh? Would love feedback — or attempts to break it.