Select two. Developers say the automated PR review flags too many non-issues, such as preferred naming styles, an intentional early return, or a local pattern the team uses deliberately. Adding "only report findings you are highly confident about" to the prompt made no measurable difference. Which two changes will most effectively reduce false positives?

Answer: A and C — Replace the confidence instruction with explicit criteria listing which categories to report (bugs, security issues) and which to skip (minor style, local conventions). / Add a few targeted examples contrasting acceptable code patterns from your codebase with genuine issues, explaining why each one is or isn't worth flagging.

Section 6 says specific categorical criteria improve precision where general instructions like "be conservative" or "only report high-confidence findings" don't (a). Few-shot examples that distinguish acceptable patterns from genuine issues reduce false positives and still let the model generalize (c). (b) is the same confidence-based wording that already failed, just repeated louder. (d) leaves the cause, unclear criteria, in place.