← Back to AI & LLMs
AI & LLMs by @sebastianffx

user-cognitive-profiles

Analyze ChatGPT conversation exports

0
Source Code

User Cognitive Profiles

🤖🤝🧠 Discover how you communicate with AI and optimize your agent interactions.

This skill analyzes your ChatGPT conversation history to identify cognitive archetypes — recurring patterns in how you think, communicate, and collaborate. Use these insights to calibrate your OpenClaw agent for more effective, personalized interactions.

Why This Matters

Human-AI communication is not one-size-fits-all. Just as you adapt your communication style between contexts (work meeting vs. casual chat), effective AI assistance requires matching your cognitive architecture.

The Problem:

  • Default AI behavior assumes a generic user
  • Your communication style varies dramatically by context (professional vs. personal)
  • Misaligned AI responses feel inefficient or frustrating

The Solution:

  • Analyze your actual conversation patterns
  • Identify your dominant cognitive archetypes
  • Configure your agent to match your communication style

Quick Start

1. Export Your ChatGPT Data

  1. Go to ChatGPT → Settings → Data Controls → Export Data
  2. Click "Export" and confirm
  3. Wait for the email (usually arrives within 24 hours)
  4. Download the ZIP file from the email link
  5. Extract it — you'll find conversations.json

2. Run the Analysis

cd /path/to/user-cognitive-profiles
python3 scripts/analyze_profile.py \
  --input ~/Downloads/chatgpt-export/conversations.json \
  --output ~/.openclaw/my-cognitive-profile.json \
  --archetypes 3

3. Apply to Your Agent

Add to your SOUL.md or AGENTS.md:

## User Cognitive Profile
<!-- Source: generated by user-cognitive-profiles skill -->
- **Primary Archetype:** Efficiency Optimizer
- **Avg Message Length:** 47 words
- **Context Switching:** High (professional vs. personal modes)
- **Key Patterns:** Prefers direct answers, values examples over theory

### Communication Calibration
- Default to concise responses
- Provide examples + theory + hands-on steps
- Watch for professional/personal mode shifts

Cognitive Archetypes

The analysis identifies archetypes based on four dimensions:

Dimension Low High
Message Length Brief commands Extended analysis
Structure Organic flow Systematic breakdown
Depth Practical focus Theoretical exploration
Tone Transactional Collaborative

Common Archetypes

🔧 Efficiency Optimizer

  • Messages: Short, direct, action-oriented
  • Wants: Quick answers, minimal explanation
  • AI Role: Tool to get things done
  • Example: "Set up email. Use pass. Go."

🏗️ Systems Architect

  • Messages: Long, structured, comprehensive
  • Wants: Deep analysis, trade-offs, strategic thinking
  • AI Role: Collaborative partner for complex problems
  • Example: 300-word technical breakdown with multiple considerations

🧭 Philosophical Explorer

  • Messages: Varies widely, questions assumptions
  • Wants: Meaning, patterns, cross-domain connections
  • AI Role: Socratic partner for insight generation
  • Example: "How does this relate to [completely different domain]?"

🎨 Creative Synthesizer

  • Messages: Connects disparate ideas, uses analogies
  • Wants: Novel combinations, pattern recognition
  • AI Role: Ideation partner and pattern mirror
  • Example: "This is like jazz improvisation..."

Customization

Define Your Own Archetypes

Create ~/.openclaw/my-archetypes.yaml:

archetypes:
  - name: "Research Mode"
    keywords:
      - "research"
      - "analyze"
      - "compare"
      - "trade-off"
    patterns:
      - long_messages
      - multiple_questions
      - citation_requests
    
  - name: "Quick Mode"
    keywords:
      - "quick"
      - "brief"
      - "simple"
      - "just"
    patterns:
      - short_messages
      - imperative_tone
      - minimal_context

Run with custom archetypes:

python3 scripts/analyze_profile.py \
  --input conversations.json \
  --archetypes-config ~/.openclaw/my-archetypes.yaml

Adjust Cluster Count

More archetypes = finer granularity, but harder to act on:

# Simple: 2-3 archetypes
python3 scripts/analyze_profile.py --archetypes 2

# Detailed: 5-7 archetypes
python3 scripts/analyze_profile.py --archetypes 5

# Complex: 10+ (for power users)
python3 scripts/analyze_profile.py --archetypes 10

Understanding the Output

Profile JSON Structure

{
  "metadata": {
    "total_conversations": 3784,
    "date_range": "2024-01-01 to 2025-01-31",
    "analysis_date": "2026-02-02"
  },
  "archetypes": [
    {
      "id": 0,
      "name": "Systems Architect",
      "confidence": 0.87,
      "metrics": {
        "avg_message_length": 382,
        "avg_response_length": 450,
        "question_ratio": 0.23,
        "code_block_ratio": 0.45
      },
      "keywords": ["architecture", "design", "trade-off", "system"],
      "sample_conversations": ["uuid-1", "uuid-2"],
      "recommendations": {
        "ai_role": "Senior Architect",
        "communication_style": "Detailed, systematic, collaborative",
        "response_length": "long",
        "structure": "hierarchical"
      }
    }
  ],
  "context_shifts": [
    {
      "trigger": "technical_keywords",
      "from_archetype": "Efficiency Optimizer",
      "to_archetype": "Systems Architect"
    }
  ],
  "insights": {
    "primary_mode": "Systems Architect",
    "context_switching": "high",
    "communication_preferences": [
      "Examples before theory",
      "Hands-on application",
      "Cross-domain analogies"
    ]
  }
}

Key Metrics Explained

Metric Description Why It Matters
avg_message_length Average words per user message Short = efficiency mode, Long = exploration mode
question_ratio % of turns that are questions High = collaborative, Low = directive
code_block_ratio % of messages with code Technical vs. conceptual focus
context_shifts Detected mode transitions Indicates multiple archetypes at play
confidence Cluster cohesion score Higher = more distinct pattern

Privacy & Security

All processing is local. The script:

  • ✅ Runs entirely on your machine
  • ✅ Never uploads data to external services
  • ✅ Stores results in your local OpenClaw workspace
  • ✅ You control what gets shared (if anything)

Recommended workflow:

  1. Export ChatGPT data
  2. Run analysis locally
  3. Review my-cognitive-profile.json
  4. Manually add relevant insights to SOUL.md
  5. (Optional) Delete the export and raw profile

Advanced Usage

Compare Profiles Over Time

Track how your communication evolves:

# January analysis
python3 scripts/analyze_profile.py \
  --input conversations_jan.json \
  --output profile_jan.json

# June analysis
python3 scripts/analyze_profile.py \
  --input conversations_jun.json \
  --output profile_jun.json

# Compare
python3 scripts/compare_profiles.py profile_jan.json profile_jun.json

Export for Other Agents

Generate a prompt snippet for Claude, GPT, or other agents:

python3 scripts/analyze_profile.py \
  --input conversations.json \
  --format prompt-snippet \
  --output agent-prompt.txt

Output:

## User Communication Profile
- Primary style: Systems Architect (detailed, analytical)
- Secondary style: Efficiency Optimizer (brief, pragmatic)
- Context switching: High (watch for mode shifts)
- Preferences: Examples + theory + hands-on steps
- Treat as: Senior technical partner, not assistant

Troubleshooting

"conversations.json not found"

The export ZIP contains multiple files. Make sure you're pointing to:

chatgpt-export/
├── conversations.json  <-- This one
├── user.json
└── ...

"No conversations detected"

Your export might be empty or corrupted. Check:

head -20 conversations.json

Should show: [{"title": "...", "messages": [...]}, ...]

"All archetypes have similar confidence"

Try adjusting the cluster count:

# Too granular
python3 scripts/analyze_profile.py --archetypes 10

# Try simpler
python3 scripts/analyze_profile.py --archetypes 3

"Analysis takes too long"

For large conversation histories (10k+ messages):

# Sample for faster analysis
python3 scripts/analyze_profile.py \
  --input conversations.json \
  --sample 1000  # Analyze random 1000 conversations

Integration with OpenClaw

Automatic Profile Loading

Add to your OpenClaw workspace AGENTS.md:

## On Session Start
1. Read `~/.openclaw/my-cognitive-profile.json` if exists
2. Adapt communication style to primary archetype
3. Watch for context shift indicators

Dynamic Mode Detection

For agents that can switch modes mid-conversation:

# Pseudocode for agent integration
def detect_mode_shift(current_message, profile):
    for shift in profile["context_shifts"]:
        if shift["trigger"] in current_message:
            return shift["to_archetype"]
    return profile["insights"]["primary_mode"]

Contributing

Have a new archetype that works well? Submit a PR with:

  1. Archetype definition in examples/
  2. Sample data (anonymized)
  3. Validation that it clusters distinctly

References

  • references/methodology.md — Technical details on clustering algorithm
  • references/archetype-taxonomy.md — Full archetype definitions
  • examples/ — Sample profiles and configurations

Built for humans who want their AI to truly understand them. 🤖🤝🧠