Use case · Developers

Prompts in review, like the rest of your code.

If a prompt decides what your product says, it needs a history, a test and an owner.

support-reply.promptv2 → v3

1section · Write the reply

−Answer the question in at most ten sentences.

+Answer the question that was actually asked, in at most six sentences.

4If you cannot resolve it, say what happens next and when.

commit Shorter reply, no promises

branch main

tag launch

committed to main as v3

A workflow that feels familiar.

Keep the prompt out of the code
Store it in Prompt Verse with {{variables}}. Your code asks for the rendered messages, not a string it pasted in.
Branch for an experiment
Try a new wording on a branch. Main keeps serving the version in use.
Prove it with an eval
A few dozen cases, a scoring method and a threshold. A change that drops the score does not get merged.
Tag what ships
Tag the version that passed. Your service fetches that tag.

Commits with messages
A reviewable history instead of “prompt_final_v7”.
Messages and JSON out
Render as chat messages, JSON, YAML, XML or Prompty.
A door for your agents
Your coding assistant can read and improve prompts over MCP.

Pre-beta
Fetching at run time
An HTTP API and an MCP server exist in the pre-beta. Check the reference for the routes before you depend on them.
Not offered
No SDK yet
There are no client libraries. You call the HTTP API or MCP directly.