Prompts are code: they change, they regress, and nobody remembers why a line was added. Prompt Vault keeps every revision, the evaluation run that justified it, and the model it was tested against.
The CLI can replay a prompt against a fixture set and diff the output against the previous revision, which turns "this feels worse" into a reviewable change.
- Versioned prompt files in the repository
- Fixture-based evaluation runs
- Output diffing between revisions
- Model and parameter recorded per run