Post by Amil Bhagat

i stopped trusting single-agent pr reviews. so i built a jury. ever since i started using claude code, i can't stop tinkering with skills. they're weirdly addictive to build. here's one i'm happy with: review-committee there are already good ai review skills out there — claude code's own /code-review spins up 5 specialized agents (bugs, git blame, CLAUDE.md compliance, etc). i've used it, it works. but each agent sees a slice — if one misses something, nothing catches it. no redundancy. so i thought — what if instead of specialization, i ran a blind jury? the idea:  → spin up 5 independent agents  → each one reviews the entire pr from all dimensions (no assigned lens)  → each returns structured findings with severity scores  → match issues across agents  → only surface what multiple agents agree on then a consensus matrix decides what makes the cut — high severity and most agents flagged it? must fix. moderate severity but only a few caught it? nit. only one agent found it? probably a hallucination, drop it. single-agent findings are noise. multi-agent agreement is signal. does it burn through a few more tokens than a single review? yes. worth it? maybe. fun and correct? yes skill linked here if you want to try it or fork it: https://lnkd.in/gNiTwRRy #ClaudeCode