AI Fluency Assessment

Use --max-sessions 6000 to capture all sessions. Many session files are subagent files with no user messages, so even with thousands of files the scan is fast.

LLM classification (default) requires the anthropic SDK and API credentials. The script auto-detects Anthropic Foundry credentials from environment variables (ANTHROPIC_FOUNDRY_API_KEY), or uses ANTHROPIC_API_KEY for direct API access. Cost: ~$2-5 per full assessment with Claude Haiku.

Output files:

• .ai-fluency/evidence.json — all extracted user messages with project metadata
• .ai-fluency/analysis.json — per-behavior match counts, sample messages, top projects

Step 2: Self-Assessment Questionnaire

Ask the user the following 6 questions for behaviors that have no observable signal in conversation data. Rate each 1-5 (1=Never, 5=Always).

Present them in conversation — don't run an interactive CLI.

Delegation:

• Q3: How often do you evaluate whether a specific AI tool is the right fit for your task before starting?
• Q4: How actively do you compare and choose between different AI platforms (e.g., Claude vs ChatGPT vs Gemini) based on task requirements?

Discernment:

• Q18: How alert are you to potential biases in AI outputs (e.g., cultural, gender, or perspective biases)?

Diligence:

• Q19: How carefully do you consider data privacy, security, and organizational policies before sharing information with AI systems?
• Q20: How aware are you of ethical implications throughout your AI interactions?
• Q24: How fully do you accept personal accountability for the final quality of your AI-assisted work?

After collecting responses, save them:

python3 -c "
import json
from pathlib import Path

responses = {3: R3, 4: R4, 18: R18, 19: R19, 20: R20, 24: R24}

evidence_path = Path('.ai-fluency/evidence.json')
with open(evidence_path) as f:
    evidence = json.load(f)
evidence['questionnaire'] = {str(k): v for k, v in responses.items()}
with open(evidence_path, 'w') as f:
    json.dump(evidence, f, indent=2, default=str)
print('Questionnaire saved')
"

Replace R3, R4, etc. with the user's actual integer ratings.

Step 3: Score All 24 Behaviors

Read .ai-fluency/analysis.json for the 18 LLM-classified behaviors and .ai-fluency/evidence.json for the questionnaire responses. Score each behavior 1-5:

Scoring criteria:

• 1 - Novice: No evidence of this behavior
• 2 - Emerging: Occasional or inconsistent evidence
• 3 - Developing: Regular evidence but room for improvement
• 4 - Proficient: Consistent, effective demonstration
• 5 - Expert: Sophisticated, nuanced mastery

For LLM-classified behaviors (B1-B2, B5-B17, B21-B23): Base scores on LLM match counts, quality of sample messages, consistency across projects, and range/variety within the behavior.

Important: High frequency alone does NOT guarantee a high score. Assess quality and range, not just quantity.

For questionnaire-only behaviors (B3, B4, B18, B19, B20, B24): Use the self-assessment rating directly.

Step 4: Generate the Visual HTML Report

Generate a self-contained HTML report at .ai-fluency/fluency-report.html following the report spec for design system, layout, and formatting specifications.

Step 5: Open the Report

open .ai-fluency/fluency-report.html

Framework Reference

See the framework reference for the full 24-behavior reference table.