Why games made by (only) LLM suck

I set out to see how far I can push the current leading LLMs in fully autonomous game development. My question was: what is stopping LLMs from producing games at the rate it already produces software and websites?

I wrote a prompt, asked Claude, Codex and Gemini separately to build a game (one-shot) based on the prompt. When something failed, I iterated on the prompt to avoid the mistake for the next run. After a day of tuning, filling holes, writing documentation for common APIs, and generating a template project in addition to the prompt, I gave up on this experiment.

The thing missing is obvious to me now: playtesting.

A human won't produce a good game either, if it can't ever play the game.

I successfully incorporated reinforcement learning in the template & test suite. This produced games which weren't 100% broken, but still incredibly confusing for a human and rarely interesting. RL helps build functioning games, but doesn't factor in other aspects (graphics, clarity, fun).

I see two ways forward for agentic playtesting:

• Pixel-based: LLM+Vision or pixel->input transformer
• Text-based: LLM text only. For games where the state and actions can be explained fully in text.

Text-based looks like the lowest-hanging fruit: the game could be fully played in a chat conversation with the LLM. After the testing, the LLM might even have insight on if the game was interesting, fun, difficult.

2026-01-30