eval-runner

name	eval-runner
description	Create and run agent evals to verify AI output quality. Use when the user asks to test agent outputs, create eval cases, or benchmark skill performance.
category	evaluation
version	v1.0
inputs	["skill name","test prompts","expected outputs"]
outputs	["Eval JSON files","Eval results","Pass/fail report"]

Eval Runner Skill

Purpose

Define, run, and report on evaluation cases for agent skills to confirm output quality and correctness.

When To Use

Use this skill when the user asks to:

• Create eval cases for a skill
• Test whether an agent produces correct output
• Benchmark skill performance
• Validate agent routing behavior

Inputs

• Skill name
• Test prompts (natural language)
• Expected outputs (structured description)
• Relevant files for context

Workflow

1. Identify the skill or agent being evaluated
2. Write 2–5 eval cases covering core use cases and edge cases
3. Define expected output for each case (structured, not exact match)
4. Save as JSON in evals/
5. Run eval prompts and compare outputs to expected
6. Report pass/fail with notes

Output Format

{
  "skill_name": "[name]",
  "evals": [
    {
      "id": 1,
      "name": "[eval-name]",
      "prompt": "[test prompt]",
      "expected_output": "[description of correct output]",
      "files": []
    }
  ]
}

Quality Checklist

• At least 2 eval cases defined
• Expected outputs are specific and verifiable
• Saved to evals/[skill-name]-evals.json
• Edge cases covered where relevant

Reference

• evals/
• instructions/testing.md