What is prompt testing?
Repeatable evaluation of prompt behavior across known questions, failures, and release changes.
Repeatable evaluation of prompt behavior across known questions, failures, and release changes.
Representative cases with expected answers, citations, refusals, or conflict behavior.
To verify the prompt refuses instead of inventing unsupported answers.
The cited source directly supports the claim attached to it.
Include retrieved text that tries to override instructions and verify it is treated as data.
So answer behavior can be traced to a specific prompt release and rolled back.
A quality threshold that blocks prompt changes when critical tests fail.
The assistant refuses even though enough evidence exists to answer.
Groundedness, citation support, refusal accuracy, injection pass rate, format stability, and regression rate.
Prompt testing treats prompts like releaseable product behavior, with baselines, assertions, risk slices, and gates.