Skill Tree
A behavioral profile of how you collaborate with Claude, drawn from your real conversation history.
Built on Anthropic’s AI Fluency Index (Feb 2026), which classified 11 observable behaviors across 9,830 conversations. The headline finding: most users iterate on Claude’s outputs — 85.7% of conversations show some refinement — but far fewer question its reasoning (15.8%) or verify its claims (8.7%). When Claude’s outputs get more polished, users tend to scrutinize them less. This is the artifact effect, and it’s what Skill Tree makes visible at the level of one collaborator.
The three observable axes (Description, Discernment, Delegation) are drawn from Dakan & Feller’s 4D AI Fluency Framework. The fourth — Diligence — happens outside the conversation and isn’t measurable from chat data.
The 11 Behaviors
Description · how you shape output
Provides examples (41%) · Specifies format (30%) · Expresses tone (23%) · Defines audience (18%)
Forgemaster ·
Conductor ·
Polymath
Discernment · how you assess reasoning
Flags context gaps (20%) · Questions reasoning (16%) · Verifies facts (9%)
Illuminator ·
Architect ·
Polymath
Delegation · how you set up the collaboration
Clarifies goals (51%) · Sets interaction style (30%) · Discusses approach (10%)
Compass ·
Conductor ·
Architect
Diligence · not observable in chat
Transparent about AI’s role · Considers sharing consequences · Deploys AI responsibly
Gateway · the most common behavior
Iterates on outputs (86%)
Catalyst
Population baselines from the AI Fluency Index (N = 9,830 conversations). Your rates appear next to these throughout the visualization.
How this is scored
- Unit of analysis: one conversation = one unit. A behavior either appeared in the session or it didn’t.
- Classifier: Claude Haiku reads each session against the 11 behavior definitions and emits present / absent with a high / medium / low confidence label and a short evidence quote.
- Your rate: sessions where the behavior was present, divided by total sessions classified.
- Baseline: per-conversation prevalence published in the AI Fluency Index (N = 9,830, Feb 2026).
- Limitations: Diligence isn’t observable in chat transcripts. Small samples (under ~20 sessions) produce noisy rates — treat the archetype as a sketch, not a verdict. The classifier is an LLM judging another LLM’s collaborator; the evidence quotes are there so you can audit any call you doubt.
Read the full design rationale→
Source on GitHub · robertnowell/ai-fluency-skill-cards