An AI code reviewer only worked after it was forced to earn the right to comment
ItaySela · reddit · 2026-07-24
An engineer says their AI code reviewer only became useful after they heavily constrained what it was allowed to say.
What changed
- Each repo keeps its own short list of real PR blockers, instead of a shared global policy.
- Every comment must point to a line in the changed diff, not pre-existing code elsewhere in the file.
- Each finding must include a required failurepath field describing exactly how the bug would break in practice.
- Anything with an empty failurepath is dropped before it reaches the team.
Result
False positives fell to near zero, and the team started reading the bot’s output again. The post argues that discipline and schema enforcement mattered more than a smarter model.
The author asks whether others let review bots speak freely and filter afterward, or gate what they are allowed to raise in the first place.
More from coding & agent
- Allie Miller: AI is Moving Towards Multiplayer, Proactive, and Autonomous Agents — alliekmiller · 2026-07-24
- Prompting agents without touching code is basically engineering management, he says — irvinebroque · 2026-07-24
- JavaScript’s New `sumPrecise()` Jokes That It Still Can’t Sum Precisely — Aizkmusic · 2026-07-24
- TRL and OpenEnv now let you train coding agents end to end with opencode — SergioPaniego · 2026-07-24
- A coding-agent user wants `/goal` to batch tasks and compact between them — mattpocockuk · 2026-07-24
- GPT 5.6 Sol and Luna reportedly built a real-time vision stack over MCP — MoonL88537 · 2026-07-24