Talking to a prompt before it goes live

Staff who tune the assistant’s prompts could already see everything a draft would change before publishing it: the text, the difference from what is live, the size of the prompt it produces, and a score. What they could not see was how the assistant would actually answer on it. The only way to find out was to publish, at which point every organisation was already talking to it.

A conversation on a draft

Each assistant now has a Talk to it page. It holds a real, streamed conversation on that assistant’s prompt with every saved draft applied, so a wording change can be tried across a few turns before anyone else sees it. The same page can run the conversation on the prompt that is live right now, which is how a change is compared against what it would replace.

The page says plainly what the conversation runs on: which source it is on, which files carry a draft, and that the conversation is bound to the demo workspace and to no other. Nothing said there is kept — no conversation, no history, no record — and reloading the page clears it.

What it cannot do

It cannot publish, and it cannot reach a real organisation’s data. A draft still reaches nobody until it is published, and the assistant’s own off switch applies here as it does everywhere else. Staff talk to the assistant as themselves, so the tools that read as a particular role are not available in the test conversation, and the page says which are.