[currere]

Since some people are complaining about the difficulty curve, you could randomly generate puzzles and rate them, and players, with Bayeselo (or similar; like chesstempo does). Then you have to use this ratings data to somehow discover how quickly to increase the difficulty of puzzles presented to the player (for example, perhaps select one with a 0.8 prob of success every time).

Actually, thinking about it some more, a player's rating should increase over time as they learn, while a puzzle's (true, latent) rating should be fixed, so you should allow for that.

Edit: love the game though, nice one!

[currere]

Though it looks like these puzzles are far from randomly generated - they're quite artfully put together.