Help us improve
Share bugs, ideas, or general feedback.
From navigator
Manages user preferences and corrections for bilateral modeling, learning from session corrections. Invoke to save, show, update, or reset your profile.
npx claudepluginhub alekspetrov/navigator --plugin navigatorHow this skill is triggered — by the user, by Claude, or both
Slash command
/navigator:nav-profileThis skill is limited to the following tools:
The summary Claude sees in its skill listing — used to decide when to auto-load this skill
Manage user preferences for bilateral modeling - enabling Claude to understand and adapt to your working style, technical preferences, and past corrections.
Learns user preferences from corrections (3+), steering patterns, periodic checkpoints, and explicit triggers to adapt Claude's behavior across sessions.
Analyzes coding sessions to detect corrections and preferences, proposing targeted updates to active Skills or CLAUDE.md. Triggered by 'learn from this session' or 'update skills'.
Captures high/medium/low confidence patterns from conversations to prevent repeating mistakes and preserve successes. Invoke proactively after corrections, praise, edge cases, or skill-heavy sessions.
Share bugs, ideas, or general feedback.
Manage user preferences for bilateral modeling - enabling Claude to understand and adapt to your working style, technical preferences, and past corrections.
Based on Riedl & Weidmann 2025 research on Human-AI Synergy:
This skill enables Claude to:
Auto-invoke when:
DO NOT invoke if:
.agent/.user-profile.json (git-ignored, session-persistent)
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
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": []
}
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}
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.
IMPORTANT: This action triggers automatically - no explicit skill invocation needed.
When to detect corrections (monitor ALL conversations):
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.
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.
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";
}
});
For explicit actions (SHOW, UPDATE, RESET): Show confirmation message.
For auto-learn (LEARN): Silent acknowledgment, periodic summaries.
{
"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"
}
]
}
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"
Preserves profile reference:
## Profile State
- Preferences loaded: ✅
- Corrections this session: {count}
- Goals active: {goal_names}
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
}
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:
Profile is:
.agent/.user-profile.json)Profile does NOT store:
Profile management succeeds when:
Good profile usage:
Avoid:
This skill enables bilateral Theory of Mind - Claude understanding you as well as you understanding Claude 🧠