How an AI Chess Trainer Works, From the Engine Up

"AI chess trainer" is stamped on half the chess apps you'll see this year, and it describes at least four different kinds of software. Some of them will genuinely teach you; some are an evaluation bar wearing a coach's whistle. The difference is easier to spot once you know what's actually inside a real one.

The short version

An AI chess trainer is software that teaches rather than merely evaluates. It sees the exact position you're looking at, checks the chess facts with an engine, explains them in ordinary language, adapts the explanation to your level, and answers follow-up questions. Every part of that sentence matters, and the rest of this article explains why.

The two components that matter

Under the hood of any trainer worth the name, two very different systems are cooperating.

The engine supplies the truth. Modern engines play far beyond any human; Stockfish, the best-known one, is free and open source. Give one a position and it returns evaluations, best lines, and tactical facts: this bishop is attacked, this pawn push wins material, this endgame is a draw with correct play. What an engine cannot do is explain. Its native output is a number like +1.7 and a wall of moves, and if you could already read those fluently you probably wouldn't be shopping for a trainer.

The language model supplies the teaching. Large language models are excellent at explaining things in plain words and adjusting to whoever's asking. But a language model on its own is a poor chess player: it generates plausible-sounding text, and in chess, plausible-sounding is often illegal. Ask a general chatbot to analyze your game and sooner or later it will move a piece that isn't there or announce a check that doesn't exist.

A real AI trainer wires the two together, engine first, words second. The engine establishes what is true in the position; the language model's only job is to translate that truth into ideas you can reuse. People in this field call it grounding, and it's the single property that separates a trainer from a chatbot with a chess vocabulary.

What happens when you ask a question

Say you played the Italian, developed with 5...Bg4, and a few moves later your b7-pawn fell. You ask your trainer: "Where did that go wrong?"

Behind the scenes, roughly this happens. The trainer reads the exact position. No ambiguity, it has the board in front of it. The engine evaluates it and the moves around it, and notices the evaluation jumped right after ...Bg4. It finds the concrete problem: with the light-squared bishop gone from the queenside, White's Qb3 hit b7 and f7 at the same time, and Black couldn't defend both. Only then does the language layer write the answer: your bishop had a defensive job on c8 that it abandoned. The pattern the coach adds on top is that early ...Bg4 in these positions always needs a plan for the Qb3 double attack.

Then you ask the follow-up a static report can't handle: "So what should I have played instead?" The trainer answers from the same ground truth (maybe ...Be7 first, or meeting Qb3 with ...Qd7), because the engine already knows the alternatives. That loop, question and grounded answer on the position you actually reached, is the training.

Why an engine alone isn't a trainer

Free engine analysis is good, and it's honest to say so: Lichess offers analysis tools at no cost, subject to its published server limits, and Chess.com's Game Review classifies moves and mistakes. If you're comfortable reading engine lines and just want your errors marked, that may be all you need.

But an eval bar teaches by implication only. It tells you that 14...Rd8 dropped half the evaluation; it doesn't tell you that you keep making the same kind of mistake, moving a defender away from a loose pawn, in game after game. Training is transfer: you improve when an idea gets named clearly enough that you recognize it next Tuesday in a different position. Numbers don't transfer. Explanations do.

The four things sold as "AI chess trainers"

When you're comparing tools, most of them fall into one of these buckets:

1. Engines with friendly interfaces. The analysis boards on the big chess sites. Accurate, free, and mute: you do the interpreting.
2. Analytics dashboards. Tools that aggregate your games into statistics (where you lose time, which phase leaks rating) and generate drills from your mistakes. Useful diagnostics, but closer to a fitness tracker than a coach.
3. Human-like sparring partners. Models trained to play the way people actually play at a given rating, rather than like an engine. The Maia research project is the well-known example of the idea. Good for realistic practice games.
4. Conversational coaches. Engine-grounded AI you can talk to about the position. This is the newest category, and the one this article has mostly described. Cody sits here.

None of these is fake, and the right pick depends on what you want. If you want realistic sparring, category three is the lane. If you want to be told why in words you can question, you want category four.

What no AI trainer can do

Honesty section. An AI trainer won't make you show up. There's no one waiting on Tuesday who'll notice you skipped. It can't read your face, your nerves, or your clock panic the way a human across the table can. And a practical note that shouldn't need saying but does: a trainer is for before and after your games. Using one during a rated online game is cheating, full stop.

Specific tools have specific limits too. Cody, for instance, handles standard chess only (no Chess960 or variants yet), runs on the web rather than as a native app, and needs an internet connection.

What this looks like in Cody, concretely

Cody is the conversational kind. It follows the board you're looking at and explains by text or live voice, drawing arrows and highlights while it talks. A strong engine evaluates every position underneath. Cody's job is the why: the plan, the threat, the pattern. You can play full games against it while it coaches, set up any position and ask about it, and solve puzzles where hints come as a nudge rather than the answer. For finished games, paste a PGN or a public Chess.com or Lichess game link and it builds a move-by-move review you can question from the board; connect a public profile to import its available game history. It adapts from complete beginner to strong club player; you set the level.

Try one before you form an opinion

The fastest way to understand an AI chess trainer is ten minutes with one, asking about a position you actually care about. Cody's live board runs in the browser on your phone or laptop with nothing to install, and it offers a short preview without an account. A free account includes unlimited text coaching. You can meet Cody as a live AI chess coach, start on the live board, or read the FAQ first if you'd rather see how it works.