When using an LLM as a judge, scoring criteria should be fixed, with one evaluator per dimension, rather than using one large prompt to evaluate all dimensions.