BotSailor
One thing that helps rules stay useful: store each rule with the review finding that created it, plus a tiny example of the bad code. When an agent breaks a rule later, you can tell whether the rule was unclear or simply ignored, and rules with no linked finding in months are good candidates to delete. It also keeps the rulebook from growing into a wall of text the model only half reads.
The contradiction audit is the part I'd steal first — rules accumulate over time, but nobody audits the rulebook. Tagging findings and counting recurrences also reads like a genuinely practical maturity metric: if the same finding keeps coming back, the rule isn't doing its job. How do you run the audit in practice — scheduled review of the whole rulebook, or triggered when a rule misfires?
I really like the idea of treating review comments as feedback about the development process, not just the code. It’s easy to fix an issue, move on, and then completely forget about it until the same thing shows up in another PR.
The part about turning recurring findings into rules feels especially practical. Over time, the harness becomes a reflection of the actual mistakes the team encounters, rather than a collection of rules someone tried to predict upfront.
I also appreciate the reminder that not everything should be automated. The goal isn’t to remove humans from reviews, but to stop spending human attention on things that can already be checked consistently. That shift toward using reviews as a way to improve the system itself is a really interesting way to think about AI-assisted development.