Ask ChatGPT for a scene, watch it get built.

The useful version of this is not “generate me a finished scene”. It is the blocking and the fiddly parts — laying out a room, arraying a hundred instances along a curve, building a three-point rig and warming the key, retiming an animation, turning a reference photo into a set. You describe the change, it happens in your tab, and you carry on working directly if it is faster to drag something than to say it.

That works because ChatGPT is handed the app's real command layer rather than a scripting sandbox: the same operations the menus and keyboard shortcuts run, with the same validation. It reads before it acts — the scene contents, the current selection, a viewport capture — and reports what actually happened afterwards, so when Studio clamps a value or refuses an operation, that is what comes back rather than a claim of success.

One thing worth knowing up front: ChatGPT reads the tool list once, when the connection is made, and Mixos widens that list as your document grows — add a cloner and the cloner commands become available. A long conversation can therefore be working from the list it was handed at the start. Pressing Refresh on the connector and starting a new chat picks up the rest; inside a conversation, asking it to search for what it needs still reaches everything.