How do you version control and manage prompts in produc...

Prompts that drive production behavior deserve the same rigor as application code, not the informal treatment of a text snippet someone edits directly in a config panel. That means storing them as structured files in Git, requiring pull-request review for changes, and tagging versions semantically — a major bump for anything that changes the output format in a breaking way, a minor bump for new capability, a patch for wording fixes — so consumers of a prompt know at a glance how disruptive a given change is.

A registry layered on top of Git storage adds what version control alone doesn’t capture well: which model each prompt version is compatible with, its evaluation scores, whether it’s deployed to staging or production, and who approved the promotion. Every change should trigger automated evaluation before it can merge, and the same infrastructure that enables that gate also enables one-click rollback and gradual A/B rollout to a fraction of traffic before a full switch.