practice

Prompt Versioning

Treating prompts as versioned, reviewed, tested artefacts rather than as strings edited in place.

llmdeploymentevaluation

A prompt is production logic. Changing one changes system behaviour in ways that are hard to predict and easy to regress, and a prompt edited directly in a console has skipped review, testing and rollback.

What treating it properly looks like: prompts in version control, changes going through review, each version identified so that every logged response records which prompt version produced it — without that, a quality regression cannot be traced to its cause — and an evaluation suite run against a change before it ships, with a pass threshold.

The other dimension is the model version. A provider updating the model underneath a stable prompt changes behaviour without any change on your side, which is why pinning model versions and re-running evals before adopting a new one is part of the same discipline.