Write the Evaluation Before Tuning the Prompt

Prompt tuning is hard to assess when the definition of “better” changes after every answer. Freeze a small acceptance set first.

For a document-routing task, define the allowed destinations and label a collection of synthetic examples. Include ambiguous inputs that should remain unresolved. Record which errors matter most, such as sending a confidential document to the wrong queue.

Run the original prompt and one revision against the same cases. Inspect changed decisions and keep examples that the revision made worse. A stronger average score can hide a serious regression in a small category. The evaluation should help you decide whether to adopt the change, not simply provide a number that celebrates the latest prompt.

OpenAI documentation.