What is an AI chess coach? What it should do, and what it shouldn't

"AI chess coach" can mean three very different things. Some apps put a friendly face on an engine: the same best move and evaluation bar, with a speech bubble on top. Others let a general-purpose chatbot talk about your position, which sounds wonderful until it tells you a knight can reach a square it cannot. And a few do the harder thing: they make the engine and the language model work together, so the advice is both correct and understandable.

Walnut Chess is built as the third kind. This article explains what that means, and gives you a checklist for judging any AI coach, ours included.

Engines are right, but they don't teach

Stockfish is far stronger than any human who has ever played. Ask it about a position and it will give you the best move, a number, and a line of moves it expects. For a learner, that is the answer to the wrong question. You don't need to know which move scores +0.35 instead of +0.20. You need to know what you should have been thinking about, which of the moves you considered were fine, and what to remember next time.

We wrote about this gap in why a coach beats a raw engine line. The short version: an engine compresses its knowledge into numbers, and someone has to unpack it.

Chatbots teach, but they aren't right

Large language models are superb at explaining things. They can adjust to your level, answer follow-up questions and describe a plan in a sentence. What they are not good at, on their own, is calculating chess. Ask a general chatbot to analyse a position from a description or a FEN and it will often miss a piece, invent a threat, or recommend an illegal move, all in the same confident tone as its correct answers.

That confidence is the real problem. A beginner cannot tell which parts of a fluent explanation are wrong, so they learn the mistakes along with the good ideas.

The combination: the engine decides, the coach explains

The fix is a division of labour. A growing number of chess coaching projects have landed on the same design: the engine does all the chess, and the language model only puts the engine's findings into words. It never has to evaluate a position itself. It reads structured facts, such as the evaluation, the best moves, the threat, the grade of the move you played, and turns them into advice.

In Walnut Chess that looks like this:

Stockfish grades every move and finds the candidate moves in each position. The coach doesn't invent grades.
The coach gets those facts with every question. When you ask "why not Nf5?", it answers from the engine's analysis of Nf5 in your actual position.
Claims about lines are checked. If you ask "what if I'd played knight to e5?", the answer comes from the engine playing it out, not from the model imagining it.
The coach's job is the human part: choosing the one idea that matters, saying it at your level, connecting it to your plan and to mistakes you've made before.

This doesn't make the coach infallible. It can still choose a less useful idea to highlight, or phrase something loosely. But it is not left to guess what is objectively happening on the board, and that is where ungrounded chatbots fail worst.

What a coach does that an engine never will

Being right about the position is necessary. It is not what makes something a coach. A coach also:

Finds out where you are. Your first visit to Walnut Chess is a short chat about your experience, with a few placement puzzles for players who already play. Beginners skip them.
Makes a plan. Where to start in lessons, which puzzles to train, how strong your opponent should be.
Guides you during play, at the moment you're stuck, not twenty moves later. See how guided play works.
Reviews with you afterwards, picking out the two or three moments that decided the game.
Remembers. Your weak themes, your openings and your recent games shape what it suggests next.
Lets you argue. "But what about this move?" is one of the best learning questions there is, and a coach answers it about your position.

A checklist for any AI chess coach

Whichever app you use, these questions separate a coach from a chatbot or a dressed-up engine:

Is it grounded? Does it say where its analysis comes from? Can it be wrong about whether a move is legal?
Does it explain in ideas? One reason in a sentence beats a ten-move variation.
What does it cost? Some apps keep much of their coaching behind a subscription. Check what you actually get without paying.
Does it let you think first? A coach that shows the answer before you've tried is training you to copy.
Does it know you? Advice should change with your level and your history, not be the same for everyone.
Does it pick moments? A good review stops where something happened, instead of commenting on every move.
Does it answer follow-ups? If you can't ask "why?", it's a report, not a coach.

What an AI coach can't do (yet)

Honesty matters here. An AI coach is patient, always available and doesn't mind the same question five times. It is not a replacement for a strong human teacher who has watched you for months, notices your nerves in a tournament, or knows when you need encouragement more than analysis. We explore that trade-off in AI coach or human coach?

What it can do, for most players, is fill the space between "I play online and hope to improve" and "I pay a coach every week". That space is where most chess players live, and it's the space Walnut Chess is built for. Which is also why it's free: a coach for players who can't justify paying for one has to cost nothing. Walnut Chess has no subscription and no premium tier, and it will stay that way. More on that.