HOW TO MAKE LLMS SHUT UP

Summary from the TLDR Founders newsletter

Greptile's AI code review bot initially generated too many low-quality comments, frustrating users. Prompting tricks and using the LLM to judge comment severity didn't solve the problem. The key insight was that comment quality standards vary by team. By clustering past comments into "good" and "bad" based on user feedback, then blocking new comments too similar to "bad" clusters, Greptile reduced noise and boosted the percentage of comments addressed by developers from 19% to 55%+.