Large language models (LLMs) are traditionally perceived as text generators. However, with a bit of ingenuity, they can be made to perform much more complex tasks. For instance, developer Andrea Ricci has found a way to teach the Claude Sonnet model to play the classic shooter DOOM. To achieve this, the game was adapted for the SCINTIX P4 platform – a single-board computer based on Raspberry Pi CM4/CM5 with ESP32-P4 and ESP32-C6 processors. The game runs on the P4 and is displayed on a MIPI DSI screen with a resolution of 1024×600. However, instead of a continuous stream of frames, the game transmits them to the Claude Sonnet model via WebSocket only step-by-step, simulating human perception. The LLM receives information about the game world and independently generates movement and shooting commands. The system demonstrates an impressive capability: Claude generally accurately describes the environment, navigates through levels, opens doors, and shoots at enemies. Special methods are used to help the model orient itself: fan-shaped depth mapping to identify walls and prevent collisions, as well as an ASCII auto-map to track traversed areas. Importantly, the LLM plays without any additional hints beyond those available to a regular player. This project shows how text models can adapt to visual perception and interaction with graphical environments, opening new horizons for the application of artificial intelligence.