Design prompts with explicit criteria to improve precision and reduce false positives

Tell the model to be conservative and report only high-confidence findings.

Confidence is self-assessed against no standard, so the instruction changes the hedging in the wording and not which findings are produced.

Precision comes from categorical criteria — the condition under which something is reportable — not from asking for more certainty.

Ask it to check that the comments are accurate.

Every reader draws the line somewhere different, so the same input yields a different set of findings each run and none of them is wrong by the instruction.

State the testable condition: flag a comment when the behaviour it claims contradicts what the code does.

Keep a noisy category running while you improve the prompt — some of its findings are real.

Developers stop reading the whole report, so the categories that were working lose their audience too.

Disable the noisy category, restore trust in the rest, and re-enable once its criteria are specific.

Name the severity levels and let the model apply them.

"High" and "medium" have no shared referent, so the same issue lands in different buckets across runs and severity stops sorting anything.

Anchor each level to a concrete example of code that belongs in it.