Yeah. I'm building a LLM game, and it works until it doesn't, then of course you can retry, but the user experience sucks, and costs levitate. Grammar guidance

Yeah. I'm building a LLM game, and it works until it doesn't, then of course you can retry, but the user experience sucks, and costs levitate. Grammar guidance helps a bit, but you still pay for retries unless you host your own model, which is another can of worms. And that's before you get into all the issues of suddenly steering the ai after a long game into providing scoring reliably, sometimes it confuses the scenario from the game proper, or hallucinates actions the user never did. I had the prototype coded in a evening. But the rest has been a month of banging rocks to get it to work reliably start to end (well I've yet to solve scoring, but still, the rest works.)