Crabhaus

What it is

Treating prompts as versioned artifacts, like code, instead of strings buried in app source. A prompt registry stores each prompt with an immutable version history, metadata, and labels such as staging or production, so a team always knows which exact prompt text produced which output.

How it works

Prompts live in a registry that assigns a new immutable version on every edit. Apps fetch a prompt by name plus a label or pinned version at runtime, so promoting or rolling back is a label move, not a redeploy. Candidate versions run against an eval set before promotion, and traces record which version served each call, making regressions diagnosable.

A concrete example

MLflow's Prompt Registry versions prompts with aliases and ties them to evals; LangSmith keeps commit-style history; promptfoo diffs two prompt versions against a test suite in CI, so a wording change that silently breaks tool-calling fails the build instead of production.

Why it matters

Prompt edits change model behavior as much as code changes, but many teams ship them without review, tests, or rollback. Versioning gives you diffs, reproducible incident debugging, honest A/B comparisons, and instant rollback; without it, the prompt that worked last week may be unrecoverable.