Built a middlegame planning trainer - please try to break it

Hi everyone, I've been building a small chess trainer around a problem I have myself: the opening ends, I have a perfectly normal position... and I'm not really sure what the plan is. Yes yes I know, I need to study my openings more... Anyway, PlanTrainer pulls your recent Lichess games and finds middlegame positions you actually reached. You write your own plan, play it out against Stockfish, and use the coach to pressure-test it. I've seen the (fair) criticism of AI chess tools that are basically Stockfish + an LLM narrating the engine output. I didn't want to build that, and spend a lot of time fine tuning, adding context, and being genuinely helpfull... LLMs are bad at chess anyway, just look at the recent fly thing... Legal moves, captures and recaptures, material outcomes and threats all get computed before the model says anything. It has to build its answer out of those facts, and if it makes a concrete tactical claim that doesn't trace back to one, the server throws the reply away and makes it try again. It's free, no signup, nothing to install. Still early, so tell me what works, what sucks, or what is confusing. https://ai-chess-coach-xi.vercel.app/ Especially interested in whether the coaching actually helps you **think about the position**, rather than just handing you moves. PS. I have a limited budget, so the AI credits can run out, sorry about that, I'll add when I can