Prompt versioning is the practice of tracking every meaningful change to prompts, model settings, variables, tools, schemas, and evaluation results in LLM applications. Without it, teams struggle to answer basic questions after incidents: what changed, who changed it, and can we roll back? The guide covers naming conventions (feature.task.version or semantic versioning), storage options, environment separation (draft/staging/canary/production), diff tracking, attaching eval suites to each version, logging prompt versions on every production request, using aliases for fast rollbacks, and building a release checklist. A practical 10-step workflow ties it all together, from creating a draft to monitoring post-release and feeding production failures back into regression test sets.
