Just a small demo of Material UI for PICO-8. The interesting thing is that the code was entirely written by GPT-5 + Claude 4.1 Opus, with me just providing them with the input about the bugs/things to change.
Obviously, those LLMs cannot create efficient PICO-8 code - there's already 6.5k tokens even though this is such a simple demo. But it's interesting how far they got since 2022. And, of course, they still cannot create nice-looking graphics or music/sfx for PICO-8, at least not directly. They might make something, but it won't look/sound good.
Even after repeated rounds of fixing bugs with the LLMs, there are some bugs left. But it's still a nice little demo in my opinion :)
Supports both the traditional input and devkit mouse (+ scroll wheel). To scroll with the controller, you can either move the cursor to the top/bottom of the screen, or move the scroll bar itself. I kept the code as-is with all the LLM comments.
just amazing that it could write this.
how many prompts did you have to give the AI?
Honestly it wasn't little, in the end probably like 30+ separate messages from me.
But most of those were for fixing little things: it can't draw icons really well, so it struggled with the sun/moon icons a lot. Also in one case a few elements were drawn off-screen, and it couldn't fix it for ~4 messages.
I think the experience with LLMs is that they currently replicate the 80/20 Pareto principle - they'll easily write the basic implementation, but they can still have issues with polishing and detail.
Here's how this demo looked after the very first message I sent to GPT-5 in Codex CLI (it made this in ~2 minutes):
As you can see, it already looked almost the same as it does now. It also had all the animations like it does now, dark mode, but it was completely hardcoded. In subsequent messages I asked the models to make it work like immediate GUI, then refactor the code, separate into the app/library parts, and more polishing stuff (scroll bar, broken layout, and so on)
To be fair, I didn't 100% rely on the models' own knowledge for this. GPT-5 does know PICO-8 quite well by itself, but I also gave them most of the manual to improve the results. I edited it to be (almost) Markdown and removed irrelevant parts (e.g. guides on using graphical sprite/sfx/etc editors). That's why GPT-5 used rrectfill in code, even though it was only added in August 2025 and GPT-5's knowledge cutoff is October 2024.
I've tried lots of other smaller things and have some half-completed projects. GPT-5 can zero-shot (which means create on the first try) lots of easy/medium games for PICO-8 without issues.
[Please log in to post a comment]
