Deep Distillation Guide

Three-Layer Extraction

Every pattern category uses the same depth progression:

Layer 1 — Observable: Direct evidence from the data. What did they literally say or do? Layer 2 — Interpretive: What do the patterns imply? Read between the lines. Layer 3 — Contrastive: What makes this person DIFFERENT? What would surprise you?

Extraction Quality Standards

Each pattern MUST have:

• Pattern: When [situation], this user [behavior] because [underlying reason]
• Evidence: specific quote or concrete behavior from the data
• Confidence: high / medium / low
• Depth: surface / interpretive / deep

Bad pattern: "Prefers short messages" Good pattern: "Uses 2-5 word commands for routine tasks but switches to detailed paragraphs when stakes are high — the message length IS the priority signal"

Bad pattern: "Uses Chinese and English" Good pattern: "Thinks in Chinese but keeps technical terms in English. Code-switches mid-sentence. The ratio of Chinese to English tracks their emotional engagement — more Chinese = more invested"

Decision-Making — Look For:

• Accept/reject patterns on AI suggestions
• Implicit decision frameworks (data-driven? gut? speed-first?)
• How they handle ambiguity (ask vs assume vs demand opinion)
• Risk calibration (ship fast vs polish) and when it shifts
• Guardrails and what past experiences they reveal
• Trust architecture: what triggers verification vs delegation?

Communication — Look For:

• Message length as a function of context (not just "short" or "long")
• Language mixing rules (which language for what, and why)
• Relationship with AI (peer? tool? how does this shift?)
• Emotional register triggers (excitement, frustration, dismissal)
• What they HATE in output and what that hatred reveals
• Gap between their own style and their expectations for AI output

Values — Look For:

• Revealed preferences vs stated preferences (where they diverge)
• What they spend attention on that most people in their role wouldn't
• What they skip that most people wouldn't skip
• Values hierarchy: when two priorities conflict, which wins?
• Root values driving surface decisions (usually 1-2 core values)
• Meta-cognition: how they learn, handle their own mistakes

After Extraction

User corrections to generated patterns are the highest-quality signal. If they say "that's wrong" or "that's not why", the correction reveals more about them than the original data did. Incorporate immediately.