OzLand: Exploring Dynamic Generative Game Design with Multi-level Runtime Authorization

Figure 1: Overview of OzLand workflow.i) At the pre-play session, the human game master constructs the scenario and game components with a generative AI assistant, including interactive scenes and NPCs, objects, and game events involved in each scene.ii) The gameplay consists of text-driven role-play interaction.The player declares their action or talk with the NPCs by entering text input, while the LLM-driven system agent progresses the game and generates the output.iii) If an unprepared component is required during runtime, the system agent suggests that the GM pause the game and generate the new component.