AgentUpdated 2026-10-24

AI Evaluation Rubric Prompt

Design scorable evaluation dimensions, weights and decision rules for model output.

#评测#Rubric#EvAL#质量DeepSeekKimiClaude Code

Prompt

Design an evaluation rubric for {{task}}.
Sample output:
{{sample}}

Include:
1. Five dimensions: accuracy, completeness, relevance, format, safety.
2. Weight and 1-5 scoring criteria for each.
3. Concrete examples per score band.
4. Automatic-fail criteria.
5. Human sampling rate.
6. Ten test cases, including edge and adversarial samples.
7. A scoring template.

Criteria must be observable; avoid subjective words like good quality.

Variables to replace

{{任务}}{{示例}}

Example output

Accuracy 5: every fact is supported by the provided sources; 4: core facts correct with one minor error.

How to use it

Give examples per score band; otherwise raters use completely different scales.

Recommended tools

Related prompts