Note by @fastfinge

I should use elevenlabs voices and sound effects to create some sort of ever-evolving audiogame. Maybe an alchemy game where a new asset only needs to be generated when a new combination is discovered? And openai figures out how the objects combine. Pass it json of the names and descriptions of objects 1 and 2, and have it return object 3, the result of combining those two things. Then pass "high quality stereo sound of combining object1 with object2 to make object3" to elevenlabs. Discovered combos and the sounds can be cached on the server meaning I only need to call the API for new ones. Just throw together something in NVGT to access the server, play sounds, and let users combine things. I guess track what the user currently has server-side too, to allow for trading, and achievements for discovering new things. But once there are thousands of objects that can all be combined with one another...I don't know how to design that database. A table with object ids and names, then another table with the recipes for each object id, I guess? And store a list of the objects each user knows how to make somewhere. Then if you want to know how to make an object, you just recurse through the two things that make up that object, the two things that make up each of those objects, etc, until you have the full recipe? Do you cache recipes or walk through it every time? Does the server do all that recursing or does the client just make lots of API calls? I guess I'd do the server part in PHP and PostgreSQL because I'm a horrible person and we're already using NVGT client side so everything is awful anyway. For gameplay purposes, the more complex an object is, the longer it takes to make. Then you can have lab upgrades. Pay for lab upgrades by selling stuff for gems? Or getting gems for discovering new stuff? Royalties when someone buys/sells an object you discovered? Pass "high quality stereo loopable ambient sound of an alchemy lab" plus descriptors of the upgrades to elevenlabs to get background sound.
This sounds like something that, if I made it, would suddenly take up five hundred gigs of ram and 8 tb of disc space because I'm an untalented hack who screwed something up. But it's a free idea for someone who's actually both more skilled than me and has more time. No, I won't be vibe coding it; that's a guaranteed way to wind up with unexpected resource utilization and unexpected API bills.