I recently used one of the heavier ChatGPT models to make a mod pack for Minecraft, fixing some of the things that prevent me from playing the game (usually I lose interest quickly after starting a new world). I had it stitch a bunch of pre-existing mods together and test that they didn’t conflict or make the game too easy. And I had it make a new mod which added custom villagers to villages who point you to destinations full of tough scary monsters guarding a prize that’s needed to beat a final boss. The villagers only give directions if you use one of the expanded food mods to cook a complex recipe for them. If you defeat the bosses scattered around the world, there’s a distant Mordor-style volcano you can then travel to and defeat the final boss. And the biomes are arranged in a spiral with the size of the biomes increasing the further you get away from spawn. I also had it leave tips and instructions around loot chests around the map to teach me how to use the mods and remind me how to build automated farms. This combination of conditions actually motivates me to play the game because there’s a real goal now and a progression that keeps the game engaging all the way through. I don’t plan on releasing it anywhere because I’ve heard a lot of anecdotes from people who made mods with AI and got yelled at for it. Still cool that I can ask for a complex custom mod pack one evening and start actively playing it the next morning.

If you make more mods using AI, I would definitely recommend utilizing one of the heavier GPT or Claude models to limit the amount of rework. I did give it full access to a laptop I have dedicated to AI agent use so it could test and iterate things without my involvement. Completely understandable to not want to give an AI model that level of control. I have to imagine that if I were running the tests myself, I would have had to hand-hold it quite a lot more than I did. Your mileage may vary.