nav-profile

Navigator Profile Skill

Manage user preferences for bilateral modeling - enabling Claude to understand and adapt to your working style, technical preferences, and past corrections.

Why This Exists (Theory of Mind)

Based on Riedl & Weidmann 2025 research on Human-AI Synergy:

• Theory of Mind (ToM) is the key differentiator in human-AI collaboration success
• Users with higher ToM achieve 23-29% performance boost
• Bilateral modeling completes the ToM loop: Claude models you, you model Claude

This skill enables Claude to:

• Remember your preferences across sessions
• Learn from corrections without you repeating them
• Adapt communication style to your level
• Build a persistent mental model of YOU

When to Invoke

Auto-invoke when:

• User says "save my preferences", "remember I like..."
• User says "update my profile", "change my preference for..."
• After detecting a correction pattern (auto-learn mode)
• User says "show my profile", "what do you know about me?"

DO NOT invoke if:

• User is creating a context marker (use nav-marker)
• User wants session-specific preferences only
• User explicitly says "just for this session"

Profile Location

.agent/.user-profile.json (git-ignored, session-persistent)

Execution Steps

Step 1: Determine Action

SHOW (viewing profile):

User: "Show my profile", "What do you remember about me?"
→ Display current profile

UPDATE (explicit preference):

User: "Remember I prefer functional style", "Save that I like concise explanations"
→ Update specific preference

LEARN (auto-detect correction):

[Internal trigger after correction detected]
→ Extract and save correction pattern

RESET (clear profile):

User: "Reset my profile", "Clear my preferences"
→ Confirm and delete profile

Step 2: Load or Initialize Profile

Check if profile exists:

if [ -f ".agent/.user-profile.json" ]; then
  echo "Profile exists"
else
  echo "No profile found, will create new"
fi

Initialize new profile (if not exists):

{
  "version": "1.0",
  "created": "{YYYY-MM-DD}",
  "last_updated": "{YYYY-MM-DD}",
  "preferences": {
    "communication": {
      "verbosity": "balanced",
      "confirmation_threshold": "high-stakes",
      "explanation_style": "examples"
    },
    "technical": {
      "preferred_frameworks": [],
      "code_style": "mixed",
      "testing_preference": "tdd"
    },
    "workflow": {
      "autonomous_commits": true,
      "auto_compact_threshold": 80,
      "marker_before_risky": true
    }
  },
  "corrections": [],
  "goals": []
}

Step 3A: Show Profile (If SHOW Action)

Display current profile:

Your Navigator Profile
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

Communication Preferences:
- Verbosity: {verbosity}
- Confirmation: {confirmation_threshold} (when to verify understanding)
- Explanations: {explanation_style}

Technical Preferences:
- Preferred frameworks: {frameworks or "none set"}
- Code style: {code_style}
- Testing: {testing_preference}

Workflow Preferences:
- Autonomous commits: {autonomous_commits}
- Auto-compact at: {auto_compact_threshold}% context
- Markers before risky changes: {marker_before_risky}

Learned Corrections ({count}):
{recent_corrections_list}

Active Goals ({count}):
{active_goals_list}

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
Last updated: {last_updated}

Step 3B: Update Profile (If UPDATE Action)

Parse preference from user input:

User: "Remember I prefer functional style"
→ Category: technical
→ Field: code_style
→ Value: functional

User: "I like concise explanations"
→ Category: communication
→ Field: verbosity
→ Value: concise

Map common expressions to profile fields:

User Says	Category	Field	Value
"concise", "brief", "short"	communication	verbosity	concise
"detailed", "thorough"	communication	verbosity	detailed
"always confirm"	communication	confirmation_threshold	always
"skip confirmations"	communication	confirmation_threshold	never
"functional style"	technical	code_style	functional
"OOP style"	technical	code_style	oop
"prefer React"	technical	preferred_frameworks	[append "react"]
"prefer Express"	technical	preferred_frameworks	[append "express"]

Update and save:

// Update specific field
profile.preferences[category][field] = value;
profile.last_updated = "{YYYY-MM-DD}";

// Write to file
Write(".agent/.user-profile.json", JSON.stringify(profile, null, 2));

Confirm update:

✅ Profile updated!

Changed: {category}.{field}
From: {old_value}
To: {new_value}

This will affect future sessions.

Step 3C: Auto-Learn Correction (If LEARN Action) [AUTO-TRIGGER]

IMPORTANT: This action triggers automatically - no explicit skill invocation needed.

When to detect corrections (monitor ALL conversations):

• User says "No, I meant...", "Actually...", "Not X, use Y"
• User repeats a correction they gave before
• User shows frustration at repeated mistake

Trigger patterns to watch for:

"No, ..." → Direct correction
"I said ..." → Repeated instruction
"Actually, ..." → Clarification
"Not X, Y" → Substitution
"Always use ..." → Rule establishment
"Never do ..." → Anti-pattern
"I prefer ..." → Preference

When detected:

User: "No, I meant plural /users not /user"
→ Correction detected: REST naming convention preference
→ Auto-save to profile (silent)

Extract correction pattern:

correction = {
  "date": "{YYYY-MM-DD}",
  "context": "{what we were doing}",
  "original": "{what I said/generated}",
  "corrected_to": "{what user wanted}",
  "pattern": "{generalized rule}",
  "confidence": "high|medium|low"
}

Add to corrections list:

profile.corrections.push(correction);

// Keep last 20 corrections (rolling window)
if (profile.corrections.length > 20) {
  profile.corrections.shift();
}

Silently acknowledge (don't interrupt flow):

[Internal log: Correction saved to profile]

Sync to Knowledge Graph (if enabled in config):

# Check if knowledge graph integration is enabled
if [ -f ".agent/knowledge/graph.json" ]; then
  # Convert correction to memory
  python3 skills/nav-graph/functions/correction_to_memory.py \
    --action convert-one \
    --correction-json '{"pattern": "{pattern}", "context": "{context}", "confidence": "{confidence}"}' \
    --graph-path .agent/knowledge/graph.json

  # [Internal log: Correction synced to knowledge graph as memory]
fi

This creates a pitfall/pattern/learning memory in the knowledge graph, making the correction available via "What do we know about X?" queries.

Periodically surface learnings (every 5 corrections):

📚 I've learned from your corrections:
- REST endpoints should use plural nouns
- You prefer functional components over class components
- TypeScript strict mode is required

These will be applied in future sessions.

Step 3D: Reset Profile (If RESET Action)

Confirm before delete:

⚠️  This will delete your Navigator profile:
- {X} saved preferences
- {Y} learned corrections
- {Z} active goals

This cannot be undone.

Delete profile? [y/N]

If confirmed:

rm .agent/.user-profile.json

Confirm deletion:

✅ Profile deleted

Future sessions will start fresh.
To rebuild, use "Save my preferences" as you work.

Step 4: Update Goals (Optional)

If user mentions a goal:

User: "I'm working on the OAuth feature"
→ Add/update goal in profile

Goal structure:

{
  "name": "oauth-feature",
  "started": "{YYYY-MM-DD}",
  "context": "OAuth implementation for user login",
  "status": "in-progress",
  "last_mentioned": "{YYYY-MM-DD}"
}

Goal cleanup (auto-archive goals not mentioned in 7 days):

// Move to completed if not mentioned recently
goals.forEach(goal => {
  if (daysSince(goal.last_mentioned) > 7) {
    goal.status = "completed-or-abandoned";
  }
});

Step 5: Confirm Action

For explicit actions (SHOW, UPDATE, RESET): Show confirmation message.

For auto-learn (LEARN): Silent acknowledgment, periodic summaries.

---

Profile Schema Reference

{
  "version": "1.0",
  "created": "2025-12-09",
  "last_updated": "2025-12-09",

  "preferences": {
    "communication": {
      "verbosity": "concise|balanced|detailed",
      "confirmation_threshold": "always|high-stakes|never",
      "explanation_style": "examples|theory|both"
    },
    "technical": {
      "preferred_frameworks": ["react", "express", "etc"],
      "code_style": "functional|oop|mixed",
      "testing_preference": "tdd|bdd|manual"
    },
    "workflow": {
      "autonomous_commits": true|false,
      "auto_compact_threshold": 70-90,
      "marker_before_risky": true|false
    }
  },

  "corrections": [
    {
      "date": "2025-12-09",
      "context": "creating API endpoint",
      "original": "Created /user endpoint",
      "corrected_to": "Should be /users (plural)",
      "pattern": "REST endpoints use plural nouns",
      "confidence": "high"
    }
  ],

  "goals": [
    {
      "name": "oauth-feature",
      "started": "2025-12-07",
      "context": "OAuth implementation for user login",
      "status": "in-progress",
      "last_mentioned": "2025-12-09"
    }
  ]
}

---

Integration with Other Skills

nav-start (Session Start)

Loads profile automatically:

### Step 3.5: Load User Profile

If `.agent/.user-profile.json` exists:
- Load preferences into context
- Apply confirmation threshold
- Note active goals
- Show: "Loaded preferences from profile"

nav-marker (Context Markers)

Preserves profile reference:

## Profile State
- Preferences loaded: ✅
- Corrections this session: {count}
- Goals active: {goal_names}

All ToM Checkpoints

Respect profile settings:

// Before showing verification checkpoint
if (profile.preferences.communication.confirmation_threshold === "never") {
  // Skip verification
} else if (profile.preferences.communication.confirmation_threshold === "high-stakes") {
  // Only verify for complex operations
}

---

Auto-Learn Triggers

Correction patterns to detect:

User Pattern	Extracted Learning
"No, I meant..."	Direct correction
"Actually, prefer..."	Preference correction
"Not X, use Y"	Substitution correction
"Always do X"	Rule establishment
"Never do Y"	Anti-pattern
"I like when you..."	Positive preference
"Stop doing X"	Negative preference

Confidence scoring:

• High: Explicit correction with reasoning
• Medium: Correction without explanation
• Low: Implicit preference from behavior

---

Privacy & Data

Profile is:

• Git-ignored (.agent/.user-profile.json)
• Local only (not synced)
• User-controlled (can delete anytime)
• Session-persistent (survives context clears)

Profile does NOT store:

• Code snippets
• File contents
• Conversation history
• Sensitive data

---

Success Criteria

Profile management succeeds when:

• Profile loads at session start
• Preferences affect behavior (verbosity, confirmations)
• Corrections persist across sessions
• Auto-learn captures patterns silently
• Goals track user's current focus
• Reset cleanly removes all data

---

Best Practices

Good profile usage:

• "Remember I prefer concise explanations" (clear preference)
• "Save that I use functional components" (specific)
• "Show my profile" (verify what's stored)

Avoid:

• Storing sensitive information
• Over-correcting (let auto-learn work)
• Resetting frequently (defeats purpose)

---

This skill enables bilateral Theory of Mind - Claude understanding you as well as you understanding Claude 🧠