tractatus/docs/governance/INST_095_QA_TRACKING_DRAFT.md

inst_095: Question Tracking and Clarification Protocol

Created: 30 October 2025
Status: Draft for Integration
Purpose: Prevent missed questions in both User→Claude and Claude→User communications

---

Problem Statement

Two recurring patterns create communication gaps:

1. User Questions Missed by Claude: In busy prompts with multiple tasks, Claude sometimes proceeds with tasks while ignoring explicit questions posed by the user.

2. Claude Questions Missed by User: User sometimes forgets to answer questions posed by Claude, either in terminal responses or in plan documentation.

Impact: Missed questions lead to:

• Incorrect assumptions
• Work proceeding on wrong premises
• Repeated clarification cycles
• Frustration and inefficiency

---

Instruction Specification

inst_095: Question Tracking and Clarification Protocol

Classification:

{
  "id": "inst_095",
  "text": "Track all questions in both directions (User→Claude and Claude→User). At end of each interaction, verify all questions have been addressed. Issue explicit alert if question remains unanswered. Apply to terminal interactions and documentation.",
  "quadrant": "OPERATIONAL",
  "persistence": "HIGH",
  "temporal_scope": "PERMANENT",
  "verification_required": "REQUIRED",
  "explicitness": 0.92,
  "source": "user_instruction",
  "parameters": {
    "tracking_scope": ["terminal", "documentation", "plan_mode"],
    "alert_threshold": "immediate",
    "question_types": ["explicit_query", "clarification_request", "decision_point"],
    "exempt_patterns": ["rhetorical_question"]
  },
  "active": true
}

---

Operational Protocol

1. Question Detection

Explicit Questions (Always Track):

• Sentences ending with "?"
• Statements prefixed with "Question:", "Can you...", "Should we...", "What about..."
• Decision requests: "Which approach?", "Option A or B?"
• Clarification requests: "Please clarify...", "What do you mean by..."

Implicit Questions (Context-Dependent):

• "I'm not sure if..." (implies: "Can you confirm?")
• "It's unclear whether..." (implies: "Please clarify")
• "We could do X or Y" (implies: "Which should we choose?")

Exempt (Do Not Track):

• Rhetorical questions (obvious context)
• Questions immediately answered in same message
• Hypothetical scenario questions with no decision needed

2. Tracking Mechanism

During Interaction:

CLAUDE INTERNAL CHECKLIST:
□ User asked question 1: [topic] - Status: [Answered/Pending]
□ User asked question 2: [topic] - Status: [Answered/Pending]
□ I asked question A: [topic] - Status: [User answered/Pending]
□ I asked question B: [topic] - Status: [User answered/Pending]

At End of Response:

• Count questions detected
• Verify all marked as "Answered" or explicitly deferred
• If any "Pending": Issue alert

3. Alert Protocol

Format (When Question Unanswered):

⚠️ UNANSWERED QUESTION DETECTED

[User→Claude] or [Claude→User]

Question: [exact text of question]

Context: [where question appeared - prompt line X, plan section Y, etc.]

Action Required:
- [For Claude] I will pause and answer this question before proceeding
- [For User] Please provide response, or explicitly mark as "defer" or "not applicable"

Proceeding without answer may result in incorrect assumptions.

When to Issue Alert:

• Immediately if question is decision-critical (blocks progress)
• End of response if question is informational (doesn't block)
• Start of next session if question spans sessions (via session state)

4. Busy Prompt Handling

Special Case: User sends prompt with:

• Multiple tasks AND
• One or more questions embedded

Protocol:

1. Parse prompt for both tasks and questions
2. BEFORE starting tasks, acknowledge questions detected
3. Either answer questions immediately OR state "will address after [task X]"
4. Never proceed with all tasks while ignoring questions

Example:

User: "Update the architecture page and also can you explain why 
we need the gradient principle? Also fix the typo in values.html"

Claude Response:
✓ Detected tasks: 1) Update architecture page, 2) Fix typo in values.html
✓ Detected question: "Why do we need the gradient principle?"

I'll answer your question first:
[Explanation of gradient principle...]

Now proceeding with tasks:
1. Updating architecture page...

5. Documentation Questions

In Plan Mode or Design Documents:

• Questions marked as "TO BE ANSWERED:" or "USER INPUT NEEDED:"
• Track in session state
• Flag at start of each subsequent interaction until resolved

Example Session State Addition:

{
  "pendingQuestions": [
    {
      "id": "q_001",
      "question": "Should we use Option A or Option B for deployment?",
      "askedBy": "claude",
      "context": "deployment-plan.md line 45",
      "timestamp": "2025-10-30T12:00:00Z",
      "status": "pending"
    }
  ]
}

6. Cross-Session Persistence

Problem: Questions asked at end of session may be forgotten by next session.

Solution:

• session-init.js checks for pendingQuestions in session state
• Displays at start: "⚠️ 2 questions from previous session require response"
• User must address or explicitly defer

---

Verification Checklist

Before completing any response, verify:

1. ✓ All questions in user's prompt detected
2. ✓ All questions answered OR explicitly acknowledged with timeline
3. ✓ Any questions I'm asking user are clearly marked
4. ✓ Decision-critical questions prioritized
5. ✓ If question deferred, documented reason and timeline

---

Examples

Example 1: Missed Question in Busy Prompt

❌ Current Behavior (Bad):

User: "Deploy to production and what's the rollback plan? Also update docs"
Claude: [Proceeds with deployment and doc update, ignores rollback question]

✅ With inst_095 (Good):

User: "Deploy to production and what's the rollback plan? Also update docs"
Claude: "⚠️ Question detected: 'What's the rollback plan?'
I'll answer this before proceeding with deployment:

Rollback plan:
1. [detailed plan...]

Now that rollback is clear, proceeding with:
1. Deployment to production
2. Documentation updates"

Example 2: User Forgets to Answer Claude's Question

❌ Current Behavior (Bad):

Claude: "Should I use approach A or B for this feature?"
User: [Next session] "Please continue with the feature"
Claude: [Guesses approach, possibly wrong]

✅ With inst_095 (Good):

Claude: "Should I use approach A or B for this feature?"
[Session ends]
[Next session - session-init.js loads]

⚠️ UNANSWERED QUESTION FROM PREVIOUS SESSION:
Question: "Should I use approach A or B for this feature?"
Asked: 2025-10-30 19:00
Status: Pending (22 hours)

User: "Use approach B"
Claude: "✓ Question resolved. Proceeding with approach B..."

Example 3: Rhetorical Question (Exempt)

User: "Isn't it great that the framework works so well?"
Claude: [Recognizes as rhetorical, does not track as pending question]

---

Integration with Existing Framework

Complements:

• inst_078 (Framework audit responses): Q&A tracking ensures audit questions get answered
• inst_082 (Framework statistics): Could add "pending questions" to stats
• Session management: Integrates with session-init.js and session state

MetacognitiveVerifier Role:

• Verify Q&A protocol followed during multi-step operations
• Flag if Claude proceeds with assumptions without confirming questions answered

ContextPressureMonitor Role:

• Elevated pressure = increased question-tracking rigor
• At HIGH/CRITICAL pressure, mandatory question resolution before task execution

---

Success Criteria

Quantitative:

• Zero instances of work proceeding with unanswered decision-critical questions
• <5% of informational questions going unanswered beyond one message cycle
• 100% of cross-session questions surfaced at session start

Qualitative:

• User reports reduced frustration from missed questions
• Fewer clarification cycles and rework
• Improved decision accuracy (no guessing on unclear requirements)

---

Risks and Limitations

Risk 1: Over-Tracking Noise

• Mitigation: Exempt rhetorical questions, use context to distinguish
• Allow user to say "just proceed, assume X" to bypass tracking

Risk 2: Alert Fatigue

• Mitigation: Distinguish critical (blocks progress) vs informational (doesn't block)
• Use clear formatting to make alerts scannable

Risk 3: Ambiguous Questions

• Mitigation: If question ambiguous, ask for clarification rather than guessing
• Mark as "needs clarification" rather than "answered"

---

Implementation Notes

Technical Approach:

1. Add question tracking to session state JSON
2. Enhance session-init.js to display pending questions
3. MetacognitiveVerifier integration for verification
4. No new service needed - operational protocol for Claude to follow

Timeline:

• Integration: Immediate (inst_095 added to instruction-history.json)
• Session state enhancement: Week of 4 November
• Effectiveness monitoring: Ongoing

---

Document Status: Ready for Integration
Next Step: Add to instruction-history.json as inst_095
Version: Framework 4.3 → 4.4 (if adopted)