Social Media Platform For Chess

[Movebookchess]

Hi everyone, my name is Jackson and I am a high school student who has been building a chess social platform called Movebook over the past six months. I wanted to share it here because this community understands the technical side of chess better than almost anyone, and I would genuinely value your feedback. The platform lets players post their games, receive AI-generated analysis explaining key moments in plain English, and follow other players through a social feed. I use the Lichess API for game imports and puzzle data, and the AI analysis is powered by Claude rather than a traditional engine, which I know will raise eyebrows in a community that lives and breathes Stockfish. I am fully aware of the limitations of LLM based analysis for strong players and I am not trying to replace engine accuracy. The goal is to make post game review accessible and understandable for players who struggle to interpret raw engine output. I would love to hear honest technical feedback from this group about where the analysis falls short and what would make something like this more credible to stronger players. movebookchess.com if you want to take a look.

[Necromancer]

A platform that explains in natural language why a mysterious engine move is good would be awesome, but I'm afraid this is not easy to do. Remember that chess is just a very deep calculation, that's why engines are almost unbeatable today. Given any position, the result is already decided, we just don't know it. Humans developed positional thinking to counter our very limited calculation capacity.

IMO using standard LLMs will lead to very poor and most of the time wrong analysis, unless the LLM is calling an engine tool behind the scenes.
If you find a way to combine engine analysis + AI natural language capabilities, it would be a win.

[phhnguyen]

I am so impressed with the work of a young man!!!

As I remember, someone posted here about his implementation somewhat similar to yours, but IMHO, it is a bit different and a bit better: for a given game, he uses a chess engine to analyse it first, then sends all information (chess positions, scores, PVs...) to an AI agent to get explanations in natural language. Clearly, the AI has more information to "talk" and is more likely to be correct.

You may search his post here and find his work on GitHub.

You don't have to copy anyone, but you may need to compare some different methods and/or improve your work.

Keep up your excellent work, Jackson

[Movebookchess]

A platform that explains in natural language why a mysterious engine move is good would be awesome, but I'm afraid this is not easy to do. Remember that chess is just a very deep calculation, that's why engines are almost unbeatable today. Given any position, the result is already decided, we just don't know it. Humans developed positional thinking to counter our very limited calculation capacity.

IMO using standard LLMs will lead to very poor and most of the time wrong analysis, unless the LLM is calling an engine tool behind the scenes.
If you find a way to combine engine analysis + AI natural language capabilities, it would be a win.

Thank you for this, it is actually really helpful feedback. You are right and this is exactly what I am working toward. The current implementation does pass engine data alongside the position to the AI so it has more to work with than just the raw PGN, but I know there is a lot of room to improve how deeply that engine information is integrated. The combination you described, engine accuracy with natural language explanation, is the exact goal. Still a long way to go but this gives me a clearer direction. I appreciate you taking the time to explain it.

[Necromancer]

In theory one could generate a huge "fen, best_move, GM_explanation" dataset and fine tune some LLM.
If the data is massive enough it will work, but this dataset does not exist. Even if you take all existing chess books and extracted this info somehow, I don't think it's enought data.

A more realistic possibility you can try is to use LLM vision capabilities on position diagrams. Again, the LLM will make wrong conclusions simply because it was not trained for that. Maybe your best bet is to write code to detect positional features and their changes on quiet positions. Tactical positions are a problem. Good luck.

[Movebookchess]

I am so impressed with the work of a young man!!!

As I remember, someone posted here about his implementation somewhat similar to yours, but IMHO, it is a bit different and a bit better: for a given game, he uses a chess engine to analyse it first, then sends all information (chess positions, scores, PVs...) to an AI agent to get explanations in natural language. Clearly, the AI has more information to "talk" and is more likely to be correct.

You may search his post here and find his work on GitHub.

You don't have to copy anyone, but you may need to compare some different methods and/or improve your work.

Keep up your excellent work, Jackson

Thank you so much, this genuinely means a lot. And thank you for pointing me to that implementation, I will look it up and study how they approached it. I am not trying to reinvent the wheel and if someone solved part of this problem better than I have I want to learn from it. I really appreciate the encouragement and the kind words. Keep up the excellent work yourself, Banksia looks fascinating.