Our story
The chess is never generated.
It's explained.
Every AI chess product on the market does roughly the same thing: pipe Stockfish output into a language model and ask it to explain. The results hallucinate, contradict the engine, or produce generic platitudes. The chess community has rightly learned to distrust "AI coaching."
We asked a different question: what if the language model never had to reason about chess at all? What if every chess judgment was already made — by Stockfish, by hand-tuned classifiers, by the same frameworks human coaches have used for decades — and the LLM's only job was the last mile: putting it into words?
Three coaching frameworks.
Decades of pedagogy.
Zero guesswork.
We didn't invent how to teach chess. Grandmaster-level coaches have been doing it brilliantly for decades. We encoded their frameworks into deterministic, measurable code that runs on every position you analyse.
Dan Heisman
Elements of Positional Evaluation
Seven structural elements of chess positions — Mobility, Vulnerability, Centre Control, Coordination, Speed, Time, Flexibility — measured per-position with cohort baselines spanning 21+ super-GMs including Carlsen, Nakamura, Fischer, Tal, and Capablanca. When Nadia says "your pieces aren't coordinating well here," that's a measured structural observation, not a language model's guess.
Jeremy Silman
How to Reassess Your Chess
Fifteen coaching emphasis frames — attack-the-target, improve-the-bad-piece, force-the-issue, exercise-patience, convert-the-advantage, and more — plus the static-vs-dynamic position classification that determines which coaching applies to which moment. This is the layer that decides what Nadia teaches at each move, not just that something happened.
Jacob Aagaard
Grandmaster Preparation: Positional Play
Three positional questions composed into a predicate library that runs on every position: Where are the weaknesses? Which is the worst-placed piece? What is your opponent's idea? These aren't prompts to a language model. They're deterministic computations whose answers are already resolved before Nadia sees the position.
How the pieces fit together.
Every move you analyse passes through three layers. By the time Nadia speaks, all the chess is already solved.
01
Stockfish evaluates
Every position, every candidate move, every principal variation. Engine-grade precision from Stockfish 18, running entirely on your device. This is the chess truth layer — evaluations, best moves, tactical lines.
02
Coaching frameworks analyse
Heisman's seven positional elements measure the structural character of the position. Silman's coaching emphasis frames determine what lesson this moment teaches. Aagaard's positional predicates compute plan-level evaluation. Fifteen hand-tuned classifiers detect patterns, tactics, weaknesses, and positional themes. All on-device. All deterministic. All drawing on the same principles human coaches have used for decades.
03
Nadia explains
A language model receives a complete coaching brief — with every chess judgment already made — and articulates it in plain English, calibrated to your level. She never evaluates positions. She never bridges gaps with chess reasoning. She never guesses. Every answer is already in the brief.
Same position. Different architecture.
This is move 9 of the Opera Game — one of the most famous chess games ever played, and our load-bearing test position throughout development. Here's what two different approaches say about it.
The position calls for Bg5 — an aggressive developing move that pins the knight and puts pressure on Black's uncastled king.
Typical LLM + Stockfish approach
"Bg5 looks natural — pin the knight, increase the pressure. But watch what happens next: your opponent fires b5, and suddenly you're being chased around the board. The engine is actually really keen on Be3 here. It's a quieter move, but it does something smart..."
Notation's coaching architecture
"Black's king is still sitting on e8. That's the whole story of the position — if the king hasn't found safety yet, you want your moves to come with tempo, like they're asking a question. Between the candidates, «Bg5» is the most 'make them respond' one."
One narrates the engine. The other teaches chess. Same position. Same engine data. Different architecture.
Test us.
We're not being cocky. We genuinely want to know if what we've built is useful — and there's always room to improve.
Send us a PGN of one of your games — ideally the full game, not just a single position, because Nadia works best with context across the whole game. We'll run it through the coaching architecture and send you back what Nadia says. Compare it to what you'd get from your current tools. Tell us honestly whether it helped.
This isn't an "AI replaces coaches" tool. It's a "can we get decades of coaching knowledge to more people who otherwise wouldn't use a chess coach" tool — and get them away from tools that make the chess up and don't actually help. We want to work with the chess community, not around it. If you're a coach, a content creator, an improver with opinions — we want to hear from you.
Version 1 of the coaching architecture. Built by one person. Genuinely open to being told what's wrong with it.
Send us a game →Two months from scratch.
Feb 2026
Project started from scratch. One developer, one idea.
17 Mar 2026
Shipped to the App Store. Core coaching, Stockfish analysis, Chessnut board support.
Mar–Apr 2026
Eight phases of coaching architecture: Heisman's seven positional elements, Silman's coaching emphasis frames, Aagaard's positional predicates. Each phase measured, tested, and verified against the Opera Game and real player data.
29 Apr 2026
Phase 8 closes. The coaching-pedagogy substrate is structurally complete. The Opera Game bias that started the investigation is resolved — reproducibly, deterministically, across all test variants.
Built in Britain by one person and a lot of coffee.
Try it yourself.
Download on the App Store — £2.99 →