llm-eval-harness

Steps

1. Define the feature's job in one sentence.
2. Write golden inputs: typical, edge, adversarial.
3. Write assertions as code (schema, length, banned phrases).
4. Add an LLM-judge step with a rubric file.
5. Wire it into CI to run on every PR touching prompts.

Done When

• [ ] Harness runs locally with one command
• [ ] Failing assertions block merge

Registry

Links

• Up: AI Engineer skills
• Siblings: model-selection-eval, prompt-regression-tests
• Frameworks: react, few-shot, risen, xml-structured, chain-of-thought
• Tests: testing guide