Engineer Your Game As an AI Harness

Making Games That LLMs Can Read

A game developer recently argued that instead of fighting with Claude and other large language models, creators should design games as systems built entirely out of code — what he called a "harness" for AI tools.

Kyle Kukshtel shared these ideas in a presentation titled "Your Game Is A Harness" at Indie Rodeo, a game development conference in Calgary. He has spent nearly 15 years working as a game programmer and said his experience shows a clear pattern: LLMs handle web development well but struggle deeply with game projects ¹.

The core problem, he explained, is structural. Games contain assets, visuals, sound, design documents and many systems that are not easily represented as code. When those non-code elements become important to how a game works, an LLM cannot reason about them and the project falls apart ¹.

Why Web Tools Work Better

Kukshtel pointed to Three.js, a popular library for building 3D graphics in the browser, as evidence for his approach. Many developers use it because it lets them represent game assets as simple geometric shapes — boxes, spheres and other primitives — that are fully described by code ¹.

By keeping everything inside the code, developers give AI tools a clear picture of the project. The model can read the code, suggest changes, generate new assets as code and update the game without guessing at hidden files or complex tool-specific formats.

He noted that Svelte and similar web stacks benefit from a huge amount of public source code online, which helps LLMs produce usable results. Games simply do not have that same volume of training data ¹.

Starting Small and Verifying

Kukshtel said his own turning point came when he used an early version of ChatGPT to build a moodboarding website called mood.site. The process involved copying code into the chat interface and reviewing the output carefully. It took roughly 20 hours over a few weeks — time he said he would have spent far longer on without the AI ¹.

Since then, he has found that tools like Claude Code can handle most web-based development tasks. But whenever he tried applying the same approach to a traditional game project, the results broke down ¹.

His advice for others is to start small, verify what the model produces and treat the AI as a powerful assistant rather than an autonomous creator. He also warned that relying too heavily on this style of prompt-driven development risks producing generic results — a concern echoed by critics who have coined the term "vibe slop" for low-effort AI-generated content ³.

The bigger point is that the game itself can be designed from the start as a harness for AI, rather than trying to force the AI into a system it was never meant to read.

Key facts

• LLMs handle web code well but struggle with games that rely on non-code assets.
• Three.js lets developers describe 3D game objects as code, making them AI-readable.
• Kukshtel built a moodboarding site with an early ChatGPT in about 20 hours.
• Critics warn that prompt-driven development can produce generic 'vibe slop'.

Sources

1. Your Game Is A Harness - k-hole — kylekukshtel.com
2. AI agent — en.wikipedia.org
3. Vibe coding — en.wikipedia.org