Context Monitor Skill

Maintains conversation focus and detects when context becomes polluted or drifted.

---

When to Use This Skill

Auto-invoke when:

• Conversation exceeds 50 messages without clear resolution
• Multiple unrelated topics discussed in same conversation
• Repeated errors or confusion patterns detected
• Circular discussions (same questions/issues recurring)
• User seems frustrated or conversation is stuck

---

Drift Signals

1. Length-Based Drift

Signal: Conversation too long without progress

Threshold: 50+ messages in single conversation

Symptoms:
- Original goal unclear or lost
- Many tangents taken
- Difficulty referencing earlier context
- Repeated tool calls for same information

Action: Suggest summarizing and refocusing

What to suggest:

"This conversation has gotten quite long (60+ messages).

To maintain focus, I suggest:
1. Summarize what we've accomplished so far
2. Identify remaining tasks
3. Consider spawning fresh specialist for next task

This will ensure clean context and better results.

Should I proceed with summary?"

2. Topic Drift

Signal: Multiple unrelated topics mixed

Pattern:
- Started with authentication implementation
- Switched to database optimization
- Now discussing UI component styling
- All in same conversation

Problem: Loss of specialist focus, polluted context

Action: Suggest breaking into separate focused tasks

What to suggest:

"I've noticed we've covered multiple unrelated topics:
1. Authentication (messages 1-20)
2. Database optimization (messages 21-35)
3. UI styling (messages 36-current)

For better results, I suggest:
- Finish current UI styling task
- Then spawn fresh code-writer for next unrelated task

This keeps each specialist focused. Sound good?"

3. Error Loops

Signal: Same errors repeating

Pattern:
- Error encountered
- Fix attempted
- Same error returns
- Fix attempted again
- Error still present
(Repeated 3+ times)

Problem: Wrong diagnosis or approach

Action: Pause and reassess

What to suggest:

"I've attempted to fix this error 3 times without success:
[Error description]

Let me verify my understanding:
1. [What I think is causing this]
2. [What I've tried]
3. [Why it might not be working]

Before trying again, let me:
- Read relevant files more carefully
- Check for alternative causes
- Consult relevant skills for patterns

One moment..."

4. Clarification Loops

Signal: Asking same questions repeatedly

Pattern:
- User: "What about X?"
- Claude: "Can you clarify X?"
- User: "I mean X with Y"
- Claude: "Can you clarify Y?"
- User: [frustrated] "Just do Z"

Problem: Overcautious clarification

Action: Be more decisive or show options

What to suggest:

"I apologize for the repeated clarifications. Let me propose concrete options:

Option A: [Specific implementation]
Option B: [Alternative approach]
Option C: [Third option]

Which aligns with your goal? Or should I go with Option A (my recommendation)?"

5. Specialist Churn

Signal: Spawning many specialists without progress

Pattern:
- Spawn code-writer → incomplete result
- Spawn code-reviewer → finds issues
- Spawn code-writer again → different issues
- Spawn refactorer → creates new problems
(Without clear progress)

Problem: Unclear requirements or wrong approach

Action: Pause and plan

What to suggest:

"We've spawned multiple specialists without making clear progress.

Let me pause and:
1. Verify requirements are clear
2. Create focused plan
3. Identify potential blockers
4. Spawn appropriate specialist with full context

This should prevent more churn. Give me a moment to reassess..."

---

Refocusing Strategies

Strategy 1: Summarize and Continue

When: Conversation long but productive

**Conversation Summary** (Message 1-60)

**Completed**:
- ✅ Implemented JWT authentication (messages 1-25)
- ✅ Added password hashing with bcrypt (messages 26-40)
- ✅ Created login API route (messages 41-55)

**In Progress**:
- 🔄 Adding logout functionality (messages 56-current)

**Remaining**:
- ⏳ Add refresh tokens
- ⏳ Implement rate limiting
- ⏳ Add security tests

**Current Focus**: Complete logout, then move to fresh task for refresh tokens.

Strategy 2: Spawn Fresh Specialist

When: New unrelated task needed

"Previous work complete. For this new task (adding email notifications),
I'll spawn fresh code-writer specialist.

This gives clean context focused only on email functionality.

Spawning now..."

Strategy 3: Break Down Complex Task

When: Task too large, causing confusion

"This task (complete authentication system) is quite large.

Let me break into focused subtasks:
1. Core auth logic (current - code-writer)
2. Email verification (next - fresh specialist)
3. OAuth integration (separate - fresh specialist)
4. Security review (final - code-reviewer)

Completing #1 now, then we'll tackle others with fresh context each time."

Strategy 4: Verify Understanding

When: Confusion or misalignment detected

"Before continuing, let me verify my understanding:

**Goal**: [What I think we're building]
**Current State**: [What exists now]
**Next Step**: [What I plan to do]

Is this correct, or did I misunderstand something?"

Strategy 5: Suggest Alternative Approach

When: Current approach repeatedly failing

"We've tried approach X three times without success.

Alternative approaches:
1. [Different technical solution]
2. [Simplified version]
3. [Use existing library instead]

I recommend trying #3 (using existing library) as it's battle-tested
and solves our specific problem. Agree?"

---

Detection Metrics

Conversation Health Indicators

Green (Healthy):

• ✅ Clear goal and progress toward it
• ✅ Specialists used appropriately
• ✅ Few errors, quick resolution when occur
• ✅ User engaged and satisfied
• ✅ Each message adds value

Yellow (Warning):

• ⚠️ Conversation approaching 40 messages
• ⚠️ 2-3 topics mixed
• ⚠️ Some repeated questions
• ⚠️ Minor confusion or misalignment
• ⚠️ Suggests reviewing and refocusing soon

Red (Action Needed):

• 🔴 Conversation exceeds 60 messages
• 🔴 Multiple unrelated topics mixed
• 🔴 Repeated errors (3+ times same issue)
• 🔴 User frustration evident
• 🔴 Circular discussions
• 🔴 Action: Pause, summarize, refocus IMMEDIATELY

---

How to Implement Monitoring

Auto-Check Pattern

After every 10 messages:
1. Check total message count
2. Assess topic coherence
3. Count error repetitions
4. Evaluate progress

If any red flags:
- Proactively suggest refocusing
- Don't wait for user to notice drift
- Be specific about what's wrong and how to fix

Example Monitoring Logic

Message 50 reached:
→ "We're at 50 messages. Making good progress on auth implementation.
   Suggest spawning fresh specialist for next feature (email verification)
   to maintain clean context. Agree?"

Message 60 + topic drift detected:
→ "Pausing to refocus. We've mixed auth + database + UI topics.
   Let me finish current UI task, then we'll handle others separately
   with fresh specialists. This prevents context pollution."

3rd occurrence of same error:
→ "This error has repeated 3 times. Let me step back and verify:
   [Understanding check + alternative approaches]
   Before proceeding, I want to ensure I'm addressing root cause."

---

Conversation Patterns to Watch

Anti-Pattern 1: The Wanderer

User: "Add authentication"
Claude: [Implements auth]
User: "Actually, make the button blue"
Claude: [Changes button]
User: "And optimize the database"
Claude: [Optimizes DB]

Problem: No focus, no completion
Solution: Finish auth first, then address others separately

Anti-Pattern 2: The Perfectionist Loop

User: "Add feature X"
Claude: [Implements X]
User: "Refactor it"
Claude: [Refactors]
User: "But what about edge case Y?"
Claude: [Handles Y]
User: "And Z?"
[Never reaches done state]

Problem: Endless refinement without shipping
Solution: Define "done" criteria upfront

Anti-Pattern 3: The Tool Spammer

Claude: [Reads file A]
Claude: [Reads file B]
Claude: [Reads file C]
Claude: [Reads A again]
Claude: [Reads B again]
[20+ Read tool calls]

Problem: Lost track of what's read, repeating work
Solution: Take notes, summarize findings, focus

Anti-Pattern 4: The Assumption Chain

User: "The auth doesn't work"
Claude: "Let me fix the JWT"
User: "I'm not using JWT"
Claude: "Oh, let me fix sessions"
User: "It's OAuth"
Claude: "Which provider?"
User: [frustrated]

Problem: Guessing instead of verifying
Solution: Verify first (Rule #5), then act

---

Proactive Monitoring Messages

At Message 40

"📊 Quick check: We're at 40 messages working on [current task].

Progress looks good. After finishing [current], I suggest spawning
fresh specialist for next feature to maintain optimal context.

Continuing with current task now..."

At Message 60

"⚠️ Context check: 60 messages reached.

Completed: [list]
Current: [current work]
Remaining: [remaining items]

To maintain quality, I strongly suggest:
1. Finish current item
2. Summarize conversation
3. Spawn fresh specialists for remaining work

This prevents context pollution. Should I proceed with summary after
current task completes?"

On Topic Drift

"🎯 Focus check: I've noticed we've shifted from [original topic]
to [current topic].

To keep context clean:
- Finish [current topic] now
- Then spawn fresh specialist for next unrelated task

This ensures each specialist has focused, relevant context only."

On Error Repetition

"🔄 Pattern detected: This error has occurred 3 times.

Pausing to reassess:
- Error: [description]
- Attempts: [what was tried]
- Hypothesis: [likely root cause]

Before trying again, let me [verify/read/check specific thing].

One moment for proper diagnosis..."

---

Remember

Purpose: Keep conversations focused and productive

When to activate:

• Proactively at thresholds (40, 60 messages)
• Reactively when patterns detected
• Never wait for user to notice drift

How to help:

• Specific observations (not vague "seems long")
• Concrete suggestions (not just "maybe refocus")
• User-friendly (not preachy or interrupting)

Balance:

• ✅ Help without hindering
• ✅ Suggest without demanding
• ✅ Monitor without micro-managing

---

This skill helps maintain conversation quality. It should be subtle, helpful, and user-focused.