A short run of real situations. Each one is a few lines from somebody's working day: a stack trace that will not reproduce, a PDF that needs its tables pulled out, a release nobody has documented. Your job is to name the skill that fits. Answer and you get the reasoning, including why the skills sitting next to it were wrong.
Reading a skill's description is easy. Choosing between two that overlap is not. Claude decides which skill to load from that description, so skills describing similar work compete for the same request, and the one that wins is not always the one you wanted. The skills here are grouped into families along exactly that axis: the one they collide on. The denser the family, the more it is worth drilling.
Skills you miss come back sooner and skills you have shown you know come back later. Recent accuracy nudges the difficulty, so a run sits at the edge of what you can do rather than comfortably below it. Finishing a session advances your streak. All of it is stored on your own device, which is why the game works offline and needs no account.
Every situation is written against a real public SKILL.md file, and none of them reproduces upstream prose: the descriptions here were written for this game. Before a scenario ships it is read blind by a separate model that has to pick the answer from the options alone. If that read comes back ambiguous, the question does not reach you, which is the only way a question with two defensible answers gets caught before you are the one arguing with it.
Would rather read before you play? Browse every Claude skill in the bank.