I've been running Claude Code daily since it shipped. Fast, productive by most measures — but a few weeks ago I realized I had no idea if I was actually getting better at working with it or just getting faster at the same narrow set of moves. Context-dump, ask for code, iterate. Repeat.
In February, Anthropic published a study measuring 11 observable collaboration behaviors across 9,830 Claude conversations — things like how often users provide context proactively, delegate open-ended tasks, or verify outputs critically. That's the AI Fluency Index. I wanted to know what my numbers looked like against that baseline, and more importantly, which of the 11 behaviors I never touch.
So I built skill-tree.
It installs as a Claude Code plugin in one command:
claude plugin marketplace add robertnowell/ai-fluency-skill-cards && claude plugin install skill-tree-ai@ai-fluency-skill-cards
From there, it finds your session files, extracts your side of the conversation, and ships the messages to a remote classifier (Claude Haiku on Fly.io) that maps your behavior against the same 11 categories from the Fluency Index. The behavior taxonomy comes from Dakan & Feller's 4D AI Fluency Framework — Description, Discernment, Delegation, and Diligence (that fourth one isn't visible in chat logs, so it's excluded). The whole pipeline — find files, extract, classify, assign archetype, synthesize narrative, render, return URL — runs in 30–60 seconds.
The output is one of seven archetype cards rendered as tarot cards with museum art. Here's a live example: skill-tree-ai.fly.dev/fixture/illuminator
The part I actually find useful is the growth quest: the behavior you scored lowest on (or never triggered), surfaced as a concrete prompt to try in your next session. It persists across sessions via a SessionStart hook, stored in ~/.skill-tree/ so it's there when you open Claude Code tomorrow.
If you're using Cursor, VS Code, or Windsurf instead, there's an MCP server on npm (skill-tree-ai) that exposes the same analysis without the plugin wrapper.
I found out I almost never delegate open-ended tasks — I pre-scope everything so tightly that I'm basically using Claude as a fast typist. That's probably fine sometimes. It was useful to see it named.
Top comments (0)