Danielle An and David Hoe use an audience-controlled quest game to explain why a working prompt-generated prototype is only the beginning. A key-art image can anchor assets, interfaces and gameplay ideas in one coherent style, while human taste and playtesting distinguish a good game from polished but interchangeable outputs.
Danielle An and David Hoe demonstrate competitive NPCs whose personalities and actions are generated by runtime language models rather than fixed scripts. They see room for personalized experiences and parallel iteration across disciplines, but explicitly describe the broader possibilities as experimental rather than a finished general-purpose world generator.
Danielle An and David Hoe highlight the operating challenges of model-driven games at platform scale. Model and prompt changes can alter behavior across creation, gameplay and distribution, making deterministic debugging assumptions insufficient. Inference costs, creator and player economics, content safety and repeatable evaluation remain open engineering problems.
Watch on YouTube
