I'd been using Claude Code daily for months and couldn't tell if I was improving or just running the same four workflows faster. Typing speed isn't skill. Neither is confidence.
Anthropologics published a study in February 2026 measuring 11 observable collaboration behaviors across 9,830 Claude conversations — things like whether you verify outputs, how you scope tasks, whether you delegate judgment or always retain it. I wanted to know what my sessions looked like against that distribution. More specifically, I wanted to know which of the 11 behaviors I never touched.
I built skill-tree-ai to answer that. It reads your Claude Code session history, runs a remote classifier (Claude Haiku on Fly.io) against the same 11-behavior taxonomy from Dakan & Feller's 4D AI Fluency Framework, assigns one of seven archetype cards, and picks a behavior you haven't tried as a growth quest for your next session. The whole 7-step orchestration — find files, extract messages, classify, assign archetype, synthesize narrative, render, return URL — takes 30–60 seconds.
The cards are rendered as tarot cards with curated museum art. You can see a live example at skill-tree-ai.fly.dev/fixture/illuminator. Each analysis gets a stable URL stored on a Fly.io volume.
Install in Claude Code:
claude plugin marketplace add robertnowell/ai-fluency-skill-cards && claude plugin install skill-tree-ai@ai-fluency-skill-cards
The growth quest persists across sessions via a SessionStart hook, written to ~/.skill-tree/ so it survives between runs. Cowork uses $CLAUDE_PLUGIN_ROOT/.user-state/ instead because its $HOME is ephemeral — learned that the hard way. There's also an MCP server (npm install skill-tree-ai) for Cursor, VS Code, and Windsurf.
The behavior I kept skipping: discernment — knowing when not to delegate. Hard to notice when you're in flow. Easier to notice when something classifies your last 40 sessions and shows you the gap.
Top comments (0)