Prompt drift describes the progressive degradation of a prompt that previously delivered good results and, over time, begins to produce worse, inconsistent, or outright incorrect responses. It is usually not due to a change in the text itself, but rather to external factors that alter the behavior of the underlying model.
The most common causes are:
- Model updates: the provider releases a new version with different behavior.
- Changes in default parameters such as temperature or available context.
- Data or use case drift, when real inputs begin to differ from those for which the prompt was designed.
It matters because it affects production systems: a stable integration can break without anyone having touched the code. For example, a prompt that extracted data in JSON format may, after an update, add explanatory text that breaks subsequent processing. To mitigate it, it is advisable to pin specific model versions, monitor output quality, and maintain a set of regression tests that detect degradation in time.