A review prompt defines severity as "critical: severe problems; major: significant problems; minor: small problems." The same unchecked null dereference is labeled major in one pull request and minor in the next. What is the most effective fix?

Answer: A — Define each severity level by its criterion plus a concrete code example of that level

Adjective-only definitions give the model no fixed reference, so classification drifts. A criterion plus a concrete code example for each level makes severity consistent. Defaulting low (b) biases results without making them consistent. Removing severity (c) loses useful triage. Confidence (d) measures certainty, not impact.