Most "personalised" chess training is just a filter

The short version

Personalised training is real when the material is generated from positions you actually reached. It is a filter when a fixed library of lessons, videos and puzzles is being sorted by a label you were assigned — "weak at tactics", "struggles in the endgame" — and served to everyone carrying that label.

Both feel personal. Only one of them can show you a position you have never seen in a puzzle set, because it came out of your game on Tuesday. The filter version converges: two 1200s with the same tag get the same content, and the content was written before either of them signed up.

Four questions that separate them

Ask these of any product, including ours. The answers are usually visible in the interface within a minute or two.

• Could this exact item have been shown to another player? If yes, it is library content. Not worthless — but not personalised.
• Does the plan change when your games change? Play five sharp Sicilians and then five London games. If the recommendations look identical, nothing is reading the games.
• Does it ever show you a position you lost, with you to move, at the move before you went wrong? That is the one thing a generic set structurally cannot do.
• Can you find out why an item was chosen? "Because you missed a back-rank idea in three games last week" is a reason. "Recommended for you" is a dropdown.

What the error data says a real plan should target

Across 217,199 engine-classified moves from 6,885 games analysed on Mated, 3.8% of moves played are outright blunders and another 4.2% are mistakes. That is roughly one real error every 13 moves, before you count inaccuracies, which run at 18.5%.

Those errors are not spread evenly. In the same set, 9.1% of middlegame moves are a mistake or worse, against 2.3% in the endgame. If a plan built "for you" is spending half its time on rook-and-pawn technique, it is either working from a rating band rather than your moves, or it has a good reason it should be telling you.

The third number is the one that changes how you train. The average move in that set gives away 94 centipawns against the engine's best, but the median is far lower. Most of your moves are fine or nearly fine. A small number of catastrophic ones carry almost all the damage. So the return on training is in killing the tail, not in shaving 15 centipawns off your average move. A system that ranks your weaknesses by total centipawn loss and one that ranks them by how often you drop a full piece will hand you two very different plans.

This also means "play more accurately" is bad advice for most players at this level. You are not losing because your moves are slightly off. You are losing because roughly one move in 13 is a genuine error and a few of those errors end the game.

A puzzle set you keep passing is not your puzzle set

Across 2,002 rated puzzle attempts on Mated, 94.3% were solved first time. We are not presenting that as a success. A solve rate that high means the set is sitting below the player, and most of the attempts are confirming something already known.

Puzzle ratings drift upward with a lot of easy solves, which feels good and teaches little. If you are getting nine out of ten first time, the difficulty dial is wrong, whoever set it. A set pitched at the edge of what you can calculate should be failing you regularly — enough that you notice, often enough to be mildly annoying.

There is a second problem that no puzzle set solves. In a puzzle, you know a tactic exists. In your game, nobody tells you. The skill of noticing that the position has become tactical is the skill that fails at move 22, and you cannot train it in an environment that announces itself.

A worked example

Say you are 1350 and you lose three games in a fortnight the same way. Each time, your opponent plays a quiet developing move, you reply with your own plan, and a knight you had defended two moves ago is now hanging because the defender moved. Not a combination. Not a pin you missed. You stopped checking what was attacked.

A filtered system looks at the engine tags, sees "hanging piece", and gives you the hanging-piece puzzle set. Those puzzles show you a position where a piece is already loose and it is your move to grab it. You will solve them. They train the wrong side of the mistake — capturing, not preventing.

A system reading your games gives you the position two ply before the error, with your clock running, and asks for your move. You either see it or you don't. When you don't, the point lands harder than any puzzle, because you already lost this game once.

You can do this yourself. It does not require a product. Open your last ten losses, find the move where the evaluation swung by two pawns or more, set up the position one move earlier on a physical board, and sit with it for three minutes without the engine open.

What personalisation genuinely cannot do for you

Opening choice is not a personalisation problem. No engine analysis of your last 40 games can tell you whether you will enjoy the King's Indian. It can tell you that you score 28% in one line, which is useful, but the fix is study or a repertoire change, and both are generic work.

Basic endgame technique is also generic. Lucena, Philidor, opposition, rook behind the passer — everyone needs the same handful of positions and they have not changed in a century. A plan that dresses these up as "your" weakness is being polite about the fact that you have not learned them yet.

And nobody has measured well how much faster personalised work actually is than a good generic programme at 800–1800. The claim is plausible and we believe it, but we are not aware of a controlled study, and you should treat anyone quoting a rating-gain figure with suspicion until they show the method.

The thing the software is actually buying you

Almost everything above can be done by hand with a board, your game archive and an engine. What automation buys is not insight. It is the removal of the twenty minutes of setup between deciding to review and actually reviewing, which is where most people's plans die.

So judge a training product on that axis rather than on whether it says the word personalised. Does it put you in front of your own bad position faster than you could get there yourself? If it takes longer than opening the game and scrolling to move 22, it has failed at the only job it had.