For the signal to noise reason, I start with Claude Code reviewing a PR. Then I selectively choose what I want to bubble up to the actual review. Often times, there's additional context not available to the model or it's just nit picky.