An application can look the same while its model or surrounding instructions change. A prompt that produced a useful result earlier may respond differently after an update.
Keep a small set of representative examples for an important recurring task. Include an ordinary case and one that tests a meaningful boundary. Review the outputs against the same expectations rather than judging each new result only by its style.
Read relevant change information where it is available and preserve the context of the comparison. The check does not need to become a large evaluation project. Its purpose is to confirm that the familiar task still works well enough for the way you use it.
Try the idea in context.
Consider repeating a few saved drafting tasks after changing the model. Compare the actual results instead of assuming the previous instruction behaves identically.
A thought for the next read.
Treat an example as part of the instruction. Its omissions and boundaries can influence a result as much as the words that describe the task.
- Keep representative task examples.
- Include a meaningful boundary case.
- Compare against consistent expectations after a change.
Follow a related question
Ask why the value is absent.
Missing is not a numberExplain the purpose of important choices.
Trust is visible in the detailsKeep learning
Related background to continue exploring this subject.
Google: prompt design strategies Google: an introduction to language models
