Rubric Prompting

Rubric prompting supplies the model with an explicit scoring rubric — named criteria, each with a defined scale and description of what each score means — so its judgments are consistent and grounded rather than a vague gut feel. It shows up in two places: guiding generation ("write to meet these criteria") and, more commonly, powering LLM-as-judge evaluation, where a model grades other outputs. Without a rubric, an LLM asked to "rate this answer 1-10" drifts: the same output scores differently across runs and the scale means nothing. A rubric pins the scale down — e.g., "Accuracy (1-5): 5 = every claim verifiable, 3 = one minor error, 1 = major factual mistake." For SaaS builders, rubric prompting is the backbone of automated quality evals and content grading: it makes scores reproducible, auditable, and comparable across prompt versions. Best practice is to score one criterion at a time, ask for a brief justification before the number, and validate the judge against a sample of human ratings so you trust its scores before scaling.