Tribunal skill puts code diffs on adversarial trial in Claude Code

hekman316 describes a recurring frustration with single-model code reviews: Claude either returns a cheerful approval or a vague list that blurs real bugs with stylistic preferences. The root cause, the post argues, is that a model trained to be agreeable will hedge even when explicitly told to be harsh — balance baked into one model produces politeness, not signal.

Tribunal addresses this by replacing the single reviewer with a pipeline of one-sided agents whose clash produces the honest signal.

Tribunal addresses this by replacing the single reviewer with a pipeline of one-sided agents whose clash produces the honest signal. A per-file hater agent attacks the diff as if written by a clueless amateur, focusing strictly on correctness, race conditions, memory leaks, edge cases, and security — no style nitpicks, and it is explicitly allowed to return nothing on a clean diff rather than inventing problems. A separate cross-module agent hunts bugs the per-file hater is blind to: changed function signatures whose callers still use the old API, return shapes that no longer match their consumers, and invariants out of sync across files. A judge agent then evaluates each accusation against the actual code, using docs and comments as evidence of intent — the opposite of the hater, who ignores them. Only findings the judge cannot defend, or concedes are genuinely weak, survive to the final output.

The result is a ranked report written to `docs/reviews/` plus a short chat summary organized by severity (critical → major → minor), each with a concrete fix. Tribunal is implemented as pure Claude sub-agents using the Agent tool, requires no external runtime or dependencies, and works in Claude Code and Claude Cowork across Python, JS/TS, Go, Rust, Java, and other languages configurable with one line. It is MIT-licensed and free, with installation requiring a single paste to fetch `SKILL.md` from the GitHub repo into `~/.claude/skills/`, after which `/tribunal` is available in any repo.

Key facts

1. 01Tribunal is a Claude skill built by hekman316 that reviews code diffs using adversarial multi-agent roles instead of a single model.
2. 02A per-file 'hater' agent attacks the diff on correctness, races, leaks, edge cases, and security — no style nitpicks — and may return nothing on a clean diff.
3. 03A cross-module agent separately hunts bugs invisible to the per-file hater, such as broken call signatures and out-of-sync invariants.
4. 04A judge agent evaluates each accusation against the actual code and docs, and only findings it cannot defend survive to the final report.
5. 05The final output is a ranked report written to `docs/reviews/`, organized by severity: critical → major → minor, each with a concrete fix.
6. 06Tribunal uses pure Claude sub-agents (the Agent tool) with no external runtime or dependencies, and works in Claude Code and Claude Cowork.
7. 07It is MIT-licensed and free; installation is a single paste to place `SKILL.md` in `~/.claude/skills/`, then `/tribunal` activates it in any repo.

Topics