In a 45-minute talk titled "Your Game Is A Harness" at Indie Rodeo in Calgary on 9/12/26, the author shared ideas on using LLMs in real game development, the kind of daily work done eight hours a day, five days a week, rather than one-shot demos. The author notes the post is entirely human-written, drawn from the talk's slides rather than a transcript.
The author's own journey began with mood.site, a moodboarding site built by blindly copy-pasting code into ChatGPT's web interface. Despite the clumsy process, the project was completed in about 2-3 weeks of nights, roughly 20 hours total, a timeline the author believes would have been far longer without an LLM. Since then, LLMs such as Claude Code and later Claude models have improved significantly, and for web-based stacks the author feels there is little an LLM cannot do.
Games, however, have proven a different story. The author's consistent takeaway is that LLMs are Bad At Games. Part of the reason, the author suggests, is training data: high-quality web projects are abundant online, while comparable game source is orders of magnitude scarcer. Games also contain many elements that are not easily represented as code, and they must handle dozens of systems beyond the text layout that web apps rely on.
The author acknowledges possible bias, having worked as a game programmer for nearly 15 years, which makes them quick to spot weak game code while being less critical of unfamiliar Svelte code. The talk's framing, that a game itself acts as a harness for AI-assisted development, aims to address these gaps in practical workflows.
Source: Hacker News · Summarized by HeadlinesBriefing