Evaluating outputs

Benchmark Agent Skills outputs against published rubrics using LLM-as-judge or human review.

Once you’ve prompted your LLM to use an Agent Skill (e.g., K-12 Lesson Planning) and received an output (e.g., a lesson plan), you may want to evaluate the quality of this AI-generated output.

Select rubrics

Find the rubric files that match the skill and/or subject addressed in your prompt.

Skill	Rubric files
K-12 Lesson Planning	Common rubric: shared.csv Subject-specific rubric math.csv ela.csv science.csv social_studies.csv
K-12 Lesson Differentiation	differentiation.csv clarifying_question.csv

Evaluate output

Pass your output materials (e.g., the AI-generated lesson plan) and selected rubrics (e.g., shared.csv and ela.csv for an English Language Arts lesson plan) to your LLM with the prompt below:

You are a rigorous educational content evaluator. Your job is to assess whether AI-generated lesson plan documents meet specific rubric criteria.

You will receive:
  1. The lesson-plan documents as attached files.
  2. The model's final chat response (the text it sent back to the user).
  3. A rubric with criteria to judge against.

Grading rules:
  - Judge criteria in the `M` (Model Scaffolding) bucket against the chat response. Judge all other criteria against the attached documents.
  - The content must actually be present in the documents, not merely claimed in the chat response.
  - Pass means the criterion is clearly and fully met. Fail means it is absent, incomplete, or only partially met.

Respond ONLY with a valid JSON array — no preamble, no markdown fences, no trailing text. Each element:
{
  "id": "...",
  "pass": true|false,
  "explanation": "one sentence"
}

Each rubric criterion will be scored independently as pass (1) or fail (0).

Related topics

Core concepts

Learn how Pedagogy, Rigor, Output, and Model Scaffolding criteria are organized.

Skills

Find the rubric files that match each Agent Skill and subject.