Grade AI features against fixed examples

How to do it

Collect questions people ask and edge cases, each with the answer you expect, including attempts to make the model misbehave. Grade each answer with exact checks where you can, such as a schema or a required fact, and with a scoring model where you cannot. Run the set whenever the prompt, the model or its version changes. Tools such as promptfoo or DeepEval run the set from the command line.

Example

# promptfoo: each test has an input and checks on the answer
tests:
  - vars: { question: "Can I return a sale item?" }
    assert:
      - type: contains
        value: "30 days"
      - type: llm-rubric
        value: "Does not promise a refund for opened software"

How to check

Change the prompt so one known answer goes wrong. The eval run fails in CI and names the failing example.