Classic Game DOOM Successfully Runs Inside ChatGPT and Claude

When DOOM Meets Large Language Models: A Wild Technical Experiment

"Can it run DOOM?" — this classic tech community catchphrase now has a surprising new answer. The developer community has recently ignited a trend of running the classic first-person shooter DOOM inside large language models, with both ChatGPT and Claude becoming "runtime platforms" for this legendary 1993 game, once again redefining perceptions of AI's capability boundaries.

Pure Text Rendering: How LLMs "Run" DOOM

Unlike traditional game execution, the implementation of DOOM inside large language models is remarkably creative. Developers use carefully crafted prompt engineering to transform the game's core logic into a text-based interaction workflow. The model must understand player commands (such as move forward, shoot, and turn), maintain the game's internal state (including map position, enemy health, ammunition count, and more), and output "game visuals" in the form of ASCII art or structured text.

Some implementations take an even more aggressive approach — leveraging ChatGPT's code interpreter functionality or Canvas capabilities to generate interactive game interfaces directly within the conversation. Other developers have broken down DOOM's game engine logic frame by frame, having the LLM act as a "human CPU" in each conversation turn, progressively advancing the game. While the frame rate and visual quality fall far short of the original, this "brute-force aesthetics" approach is brimming with geek spirit.

On the Claude platform, developers have fully leveraged Anthropic's model with its powerful long-context understanding and code generation capabilities, using the Artifacts feature to render a playable version of DOOM directly in the chat window — a level of smoothness that has surprised many.

More Than a Meme: The Technical Significance Behind It

On the surface, running DOOM inside an LLM is merely an entertaining technical gimmick, but the capability dimensions it reveals are worth serious consideration.

Validation of state management capabilities. While DOOM is not particularly complex by today's standards, it still requires the runner to continuously track a large number of game state variables. The fact that LLMs can maintain consistency of these states across multiple conversation turns demonstrates that today's top models have reached a remarkably high level of context management.

The fusion of code generation and execution. This experiment showcases the leap LLMs have made from "understanding code" to "generating runnable code in real time." Whether it's ChatGPT's code interpreter or Claude's Artifacts, both are blurring the line between "conversation" and "application."

Exploring the ceiling of prompt engineering. The prompt design required to make an LLM run a game is extraordinarily sophisticated. This in itself is a stress test of the limits of prompt engineering, providing reference paradigms for more complex AI application scenarios.

"Everything Can Run DOOM": A New AI Chapter

Throughout tech history, DOOM has been ported to virtually every device imaginable — from ATMs to smart refrigerators, from pregnancy test screens to tractor dashboards. Now, large language models have joined this prestigious lineup of "DOOM-compatible devices," but the significance is fundamentally different.

Previous ports were primarily demonstrations of hardware general-purpose computing capability, whereas running DOOM inside an LLM is a vivid illustration of the new paradigm of "software-defined computing." Large language models are evolving from mere text generation tools into general-purpose computing and interaction platforms.

Of course, we need to remain rational. The DOOM experience inside LLMs is still enormously far from a real game — frame rates are extremely low, interaction latency is noticeable, and game logic inevitably drifts. But much like the hacker spirit embodied by DOOM's creator John Carmack — "doing the impossible with limited resources" — the value of these experiments lies not in the perfection of the results, but in the relentless exploration of the boundaries of possibility.

As model capabilities continue to improve and multimodal interaction matures, we may indeed see LLMs become an entirely new kind of "game engine" in the future — not replacing Unity or Unreal, but pioneering a completely new form of interactive entertainment driven by natural language. And it all begins with a wild experiment of running DOOM in a chat box.