LLMs can analyze text-based Minecraft defenses to generate adversarial enemy waves. The author built a mod where an LLM interprets a fort's layout (e.g., a moat) and creates threats to exploit it, like sponge-carrying zombies draining water. This contrasts with traditional methods like Left 4 Dead’s director, which uses scalar metrics (stress or credits) rather than contextual understanding.

The LLM’s approach requires raw data input—no explicit labels for features like water or walls. For example, the model parsed an obsidian-boxed Core surrounded by water and obsidian to devise a wave targeting the moat. The author noted that scalar systems lack this granularity, as they can’t "see" strategic elements like a moat’s role in defense.

A key test involved a bug in Risk of Rain 2, where excessive credits caused the director to spawn no enemies. The LLM’s ability to adapt dynamically—without predefined rules—shows potential for emergent difficulty scaling. The experiment highlights whether LLMs can replicate human-like strategic analysis in game design.

The project emphasized raw data interpretation. The LLM received grid-based text descriptions of the fort, requiring it to deduce features like moats or choke points. This mirrors how human players strategize but automates adversarial thinking. The result? A system that could counter defenses creatively, such as sending creepers after a water-sealed Core.