PromptingIntermediateClaude API
In reviews, define what to flag and what not to flag
A categorical criterion cuts false positives far more than a general instruction to be careful.
"Only report high-confidence findings" is read differently every time. Define the exact cases that do get flagged and those that do not, and give each severity level a code example. False positives in one category make the team stop trusting the precise categories too.
Example
Flag a comment as problematic ONLY if: 1. It describes behavior that contradicts what the code does. 2. It references a function or variable that does not exist. 3. It is a TODO/FIXME for a bug that is already fixed. Do NOT flag: comments outdated only in style, minor language inaccuracies, or the absence of comments.