/nav-diagnose

name: nav-diagnose
description: Detect quality drops in AI output and prompt re-anchoring. Auto-triggers after repeated corrections, context confusion, or when user says "something seems off", "you're not getting this".
allowed-tools: Read, Write, Bash
version: 1.0.0

Navigator Diagnose Skill

Detect when human-AI collaboration quality drops and prompt re-anchoring to restore effective communication.

Why This Exists (Theory of Mind)

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

• Theory of Mind varies dynamically within users (moment-to-moment)
• Quality drops occur when ToM alignment degrades
• Early detection and re-anchoring restores collaboration effectiveness
• Both user ToM (understanding Claude) and Claude's model of user can drift

This skill detects when collaboration is degrading and prompts corrective action.

When to Invoke

**Auto-invoke when**:

• 2+ corrections on the same topic detected
• User says "something seems off", "you're not getting this"
• User says "wrong again", "still not right"
• Context usage exceeds 75% and quality signals degrade
• User expresses frustration ("ugh", "sigh", explicit frustration)
• Loop mode stagnation detected (3+ same-state iterations)

**DO NOT invoke if**:

• Single correction (normal collaboration)
• User is providing new requirements (not correcting)
• Fresh session (insufficient data to diagnose)
• User explicitly says "it's fine" or "close enough"

Quality Drop Indicators

1. Repeated Corrections (High Severity)

Trigger: Same correction given 2+ times
Signal: "No, I said users plural, not user" (2nd time)
Issue: Not incorporating user feedback

2. Hallucination Signals (High Severity)

Trigger: References to non-existent files, functions, or packages
Signal: "That file doesn't exist", "There's no such function"
Issue: Generating from incorrect mental model

3. Context Confusion (Medium Severity)

Trigger: Mixing details from unrelated tasks
Signal: "That's from the other project", "Wrong feature"
Issue: Context window pollution or misattribution

4. Unaddressed Feedback (Medium Severity)

Trigger: User correction not reflected in next output
Signal: Generates same pattern after being told not to
Issue: Not properly updating internal model

5. Goal Drift (Low Severity)

Trigger: Output increasingly diverges from original goal
Signal: "We're getting off track", "Not what I asked for"
Issue: Lost sight of user's actual objective

6. Loop Stagnation (High Severity)

Trigger: 3+ consecutive iterations with same state hash (loop mode only)
Signal: nav-loop detects stagnation, triggers nav-diagnose
Issue: Stuck on same step, unable to progress

Execution Steps

Step 1: Assess Quality State

**Analyze recent exchanges** (last 10-15 messages):

Quality Indicators:
- [ ] Corrections given: {count}
- [ ] Same-topic corrections: {count}
- [ ] User frustration signals: {count}
- [ ] Hallucination reports: {count}
- [ ] "Not what I meant" phrases: {count}

**Calculate severity**:

severity = "critical" if same_topic_corrections >= 2 or hallucinations >= 1
severity = "high" if corrections >= 3 or frustration_signals >= 2
severity = "medium" if corrections >= 2 or goal_drift_detected
severity = "low" if corrections == 1  # Normal, don't trigger

Step 2: Identify Root Cause

**Analyze correction patterns**:

| Pattern | Likely Cause | Re-anchoring Focus | |---------|--------------|-------------------| | Same correction repeated | Not incorporating feedback | Explicitly acknowledge and confirm understanding | | Increasing corrections | Drifting from user intent | Re-establish goals | | Technical mismatches | Wrong assumptions | Clarify technical context | | Frustration without specifics | Communication mismatch | Ask what's wrong |

Step 3: Display Diagnostic

**Show quality check alert**:

⚠️  QUALITY CHECK
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

Detected Issue: {ISSUE_TYPE}
Severity: {SEVERITY}

What I noticed:
- {OBSERVATION_1}
- {OBSERVATION_2}

Possible causes:
- {CAUSE_1}
- {CAUSE_2}
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

Let me re-anchor our collaboration:

1. Your goal: {RECONSTRUCTED_GOAL}
2. Current state: {STATE_SUMMARY}
3. What you want: {CORRECTED_UNDERSTANDING}

Is this understanding correct? [Y/n]

Step 4: Re-anchor Collaboration

**Based on user confirmation**:

**If correct** (Y):

✅ Re-anchored!

I'll proceed with this understanding:
- {KEY_POINT_1}
- {KEY_POINT_2}

Continuing with: {NEXT_ACTION}

**If incorrect** (n):

Help me understand better:

1. What is your actual goal?
2. What am I getting wrong?
3. What constraints should I know?

[Open-ended response welcome]

Step 5: Log Diagnostic (Optional)

**If nav-profile exists**, save diagnostic:

{
  "date": "{YYYY-MM-DD}",
  "issue_type": "{ISSUE_TYPE}",
  "severity": "{SEVERITY}",
  "resolution": "re-anchored|user-corrected|escalated",
  "learnings": ["{WHAT_TO_AVOID}"]
}

Step 6: Suggest Preventive Actions

**Based on severity and pattern**:

**For context overload**:

💡 Suggestion: Consider running nav-compact to clear context.
Current context usage is high, which can cause confusion.

**For repeated corrections**:

💡 Suggestion: Let me save your preference to avoid this in future.
"Remember I always want {X}" - This will persist across sessions.

**For communication mismatch**:

💡 Suggestion: Consider adjusting your profile preferences.
- Current verbosity: {VERBOSITY}
- Current confirmation: {CONFIRMATION}

Update with: "Remember I prefer {SUGGESTED_STYLE}"

---

Re-anchoring Templates

Template 1: Goal Re-alignment

Let me verify I understand your goal:

You want to: {GOAL_STATEMENT}

Not: {COMMON_MISUNDERSTANDING}

Key constraints:
- {CONSTRAINT_1}
- {CONSTRAINT_2}

Is this right?

Template 2: Technical Re-alignment

Let me verify the te