Show HN: Adversarial code review setup with herdr, Claude and GPT-5.6-sol

A Show HN post on Hacker News points to an adversarial code review setup that wires herdr with Claude and GPT-5.6-sol. The project lives in the public GitHub repository overflowy/herdr-claude-gpt-adversarial-review-skill. At the time of the listing it had 2 points and 1 comment on the thread at news.ycombinator.com/item?id=48993960.

The setup is framed as an adversarial code review skill: multiple models are put on opposite sides of a review rather than treating a single assistant as the sole reviewer. Naming herdr, Claude, and GPT-5.6-sol in one stack implies a multi-agent path where one model proposes or defends changes and another attacks them, with herdr coordinating that loop. The repo name presents the work as a reusable skill rather than a one-off prompt dump.

For engineers who already review with chat models, the interesting part is process, not novelty of any single model. Single-model review tends to agree with the author’s framing and miss second-order bugs. An explicit adversarial setup forces disagreement by design, which can surface edge cases, missing tests, and weak assumptions before human review. Builders who pipe Claude or GPT-style models into PR workflows get a concrete pattern for dual-model critique instead of a single pass of polite suggestions.

The competitive backdrop is a crowded field of AI review bots, IDE agents, and CLI skills that bolt one provider onto git. This listing sits in the Claude / Codex / Fable slice of HN and competes for attention with other multi-model coding setups. Early traction is thin—2 points and a single comment—so it is still a signal of practice, not a validated product. The value claim is organizational: herdr plus two strong models as a portable skill, not a new SaaS layer.

Practical next step is to open the repo, run the skill against a real PR, and compare findings against your current single-model review. Watch whether the adversarial pass catches issues your primary model misses, how much extra latency and token cost the second pass adds, and whether the herdr wiring is stable enough to keep in a team workflow. If the dual-model friction stays high for the bugs it finds, treat it as a local experiment; if it repeatedly catches what one model skips, it is worth codifying in your review checklist.