Why drills built from your own mistakes beat random puzzles

What a random puzzle actually teaches

Puzzles are good at one thing and quietly bad at another. They are good at calculation: given that something is here, find it. They are bad at recognition, because the puzzle has already told you the answer exists. You sit down knowing there is a win, so you look until you find it - and that is not the situation you are in during a game.

In a real game nobody says "there is a tactic here". The move before a blunder looks exactly like the move before a quiet position, which is why players with excellent puzzle ratings still hang pieces. The skill that failed is not calculation. It is the trigger that should have started the calculation.

The positions you never noticed are the ones you need

Consider what a puzzle set can never contain: the positions where you did not see anything. Those are exactly the positions your own games are full of. They are also the only ones that can show you what your blind spot looks like from the inside - what the position looked like at the moment you looked past it.

A drill built from your own game restores that moment. Same position, same side to move, same nothing-obvious-happening surface. The difference is that this time you have to find what you missed, and the memory of having missed it does much of the teaching.

What a good own-game drill keeps

It keeps the position exactly as it stood, your colour, and the move that was actually available. It does not tidy the position up and it does not announce the theme, because either would give away the thing you failed at. It also has to be gradeable: there must be a right answer a machine can check, or the drill becomes a discussion.

Mikhail generates these from the counted errors in your reviewed games - a position you misplayed comes back as a drill, and only a fresh one is served, so the same mistake is not recycled at you until it becomes a memory test. They are the minority of what you are served, on purpose: the server caps own-game positions at one in every ten training positions, and the rest come from the rated bank the trainer picks for your level.

Why the score has to be hard to move

Any training system that reports progress will be gamed, usually by accident. If revealing the answer still counted, the number would drift upward while your play stood still, and the record would slowly become a comfort object instead of a measurement.

So mastery rises only on a clean, unassisted solve. Revealing an answer is recorded and never rewarded, and an assisted attempt counts as an attempt without touching the streak. The result is a number that moves slowly and means something when it moves, which is the only kind worth keeping.

Where random puzzles still earn their place

Volume. You will not misplay enough positions in a week to fill a training session, and calculation does need repetition. A rating-matched tactics bank is the right filler: it keeps the hands warm between the drills that carry information about you.

So the real mix is mostly bank. At most one drill in ten is rebuilt from your own games - a cap the server enforces over every ten positions it serves - and the other nine come from the rated bank, chosen by the trainer for the skill you are currently weakest at. That ratio is deliberate: your own mistakes are scarce, and spending them all in one sitting turns a month of material into ten minutes.

How to use this if you train somewhere else

The method does not depend on this site. After each review, save the two or three positions where you went wrong into a file or a study, with the side to move and nothing else - no theme, no title, no hint. Come back to them a week later and solve them cold. If you can build only one training habit, build that one.