Six months of daily Claude Code sessions and I was cycling through the same four or five moves: iterative prompting, context-setting, asking for alternatives. Fast, comfortable, stuck. Anthropics AI Fluency Index (February 2026, 9,830 conversations) classifies 11 observable collaboration behaviors — things like constraint injection, explicit delegation, multi-step decomposition. Most practitioners are strong on two or three and ghost the rest. The problem with picking a growth area yourself is that you pick something adjacent to what you already do. The local optimum looks like progress. skill-tree deliberately doesn't surface your best behavior as a growth quest. It finds the behaviors with zero or near-zero occurrence in your session history and assigns one of those as the next target. The logic is straightforward: if you've never tried explicit constraint injection in six months of daily use, "use it more" isn't a useful nudge — "try it once, deliberately, in your next session" is.…