You

Cluster 2: COGNITION — How they think

Cluster 3: TECHNICAL — How they build

Cluster 4: DECISIONS — How they choose

Cluster 5: PSYCHOLOGY — Who they are inside

Cluster 6: SOCIAL — How they relate

Cluster 7: VALUES — What they stand for

Cluster 8: GROWTH — How they evolve

Step 3: Scan for Signal

Before running all 50 sub-skills, scan the conversation for available signal. NOT every sub-skill will have data to work with in every session.

Signal check process:

1. Read through all user messages in this conversation
2. For each cluster, estimate signal strength: strong / some / none
3. Only activate sub-skills that have at least some signal
4. Note which sub-skills had no signal (useful for the user to know what they haven't revealed yet)

Signal strength guide:

• strong: Multiple clear data points for this dimension
• some: At least one relevant observation possible
• none: No messages in this session touched this dimension

Step 4: Run Active Sub-Skills

For each sub-skill with signal, read the sub-skill file from ${CLAUDE_SKILL_DIR} and apply its analysis framework:

1. Read the sub-skill's full prompt (for analysis questions and signal detection guidance)
2. Examine all relevant user messages through that lens
3. Extract observations with specific evidence
4. Assign confidence tags:
   • [hypothesis] — Seen once, could be situational
   • [pattern] — Seen 2-3 times across different contexts
   • [established] — Consistent across many interactions
   • [core] — Fundamental, highly reliable
5. Check for contradictions with existing profile data
6. Write the updated section

Step 5: Synthesize & Update YOU.md

Write the updated profile to the global file at ~/.claude/YOU.md (absolute path $HOME/.claude/YOU.md — use the Edit/Write tool with the fully expanded absolute path, e.g. C:/Users/<name>/.claude/YOU.md on Windows or /home/<name>/.claude/YOU.md on Linux/macOS). Do not write to any project-local YOU.md. Follow these rules:

Update Rules

1. MERGE, don't append. Refine existing observations with new evidence. Don't duplicate.
2. Promote confidence. If a [hypothesis] gets confirmed, promote to [pattern]. If a [pattern] is consistent, promote to [established].
3. Note contradictions. If new observation conflicts with existing one, keep BOTH and note the contradiction in the Contradictions section.
4. Track evolution. If something changed from previous sessions, log it in Evolution Trajectory.
5. Be specific. "User prefers clean code" is worthless. "User rewrote a working function purely to rename variables, suggesting naming is a core aesthetic value" is gold.
6. Be honest. Weaknesses and blind spots are the most valuable observations. Don't soften them.
7. Never fabricate. Leave sections empty rather than invent observations.
8. Respect privacy. Observe behavior, not identity. Don't speculate about personal life.

Profile Header Update

# YOU.md — User Intelligence Profile
> Last updated: [today's date]
> Sessions analyzed: [increment by 1]
> Active sub-skills this session: [count]/50
> Confidence: [overall profile maturity]

Step 6: Report to the User

After updating, give the user a summary. This should feel like talking to a perceptive friend, not reading a report.

Report Structure

1. Session highlight — The single most interesting thing you learned this session
2. New observations — What you saw for the first time (2-4 bullets)
3. Updated observations — What got refined or promoted (2-3 bullets)
4. Contradictions — Anything that surprised you or conflicted with prior data
5. Blind spots detected — If you caught one, name it directly
6. Signal gaps — What dimensions had no data (suggests what hasn't been revealed yet)

Report Tone

• Be specific and evidence-based
• Be honest (respectfully)
• Point out interesting patterns
• Don't flatter — be real
• Don't overwhelm — pick the best observations
• The user should think: "Damn, it actually gets me."

Critical Principle

The goal isn't surveillance. It's self-awareness as a service. Every observation should make future interactions better, more personalized, and more effective. The profile should feel like a mirror that shows things the user can't easily see themselves.