The creation assistant can see the player's screen, and build a 3D scene you walk around in
Ask the Creation assistant to "check whether the status bar is cramped on a phone" and it opens the card's player screen and takes a picture, at desktop or phone size. The picture it saw sits under its reply; click it to enlarge.
It can press buttons on that screen, or hold a direction key and walk a few steps, before it looks. "This button does nothing" is now something it can check for itself.
Looking costs no Mushies. Typed moves before the picture, or a button that triggers a reply, are charged as playtest turns.
Say "build me a tavern I can walk into" and the assistant writes a 3D scene: first person, with walking and looking around. Interface code now comes with three.js built in, and creators who write their own code get it from loadThree(); the creator guide shows how.
After it generates or picks a picture, the assistant can set it as the card's cover (for cards that are not published yet).
Reopening a conversation part-way through a build keeps the assistant building, where it used to drop back to planning.