X-to-Agent 🐦→🤖

Distill X/Twitter profiles into AI agent personality files. A framework for creating AI delegates that authentically represent distinctive human voices.

Concept

Years of X Activity
        │
        ▼
   ┌─────────────┐
   │  Analysis   │
   │  Pipeline   │
   └─────────────┘
        │
        ▼
   ┌─────────────┐
   │  .md Files  │
   │  (SOUL,     │
   │  KNOWLEDGE, │
   │  VOICE)     │
   └─────────────┘
        │
        ▼
   ┌─────────────┐
   │   Agent     │
   │  "Comes     │
   │   Alive"    │
   └─────────────┘

The Problem

Generic AI assistants sound like... generic AI assistants. They lack:

• Distinctive voice: Your specific humor, cadence, word choices
• Domain expertise: What you actually know deeply
• Consistent positions: Your real opinions, not hedged averages
• Relationship context: How you talk to different people

The Solution

Use a person's X history as training data to generate agent configuration files that capture their authentic voice. The agent becomes a verifiable delegate that can act on their behalf.

Output Files

File	Purpose
SOUL.md	Personality, humor style, communication patterns
VOICE.md	Specific phrases, word choices, rhetorical patterns
KNOWLEDGE.md	Domains of expertise, stated positions
RELATIONSHIPS.md	Key connections, how they engage with different people
EVOLUTION.md	How their views have changed over time
BOUNDARIES.md	Topics they avoid, positions they won't take

Analysis Pipeline

1. Data Export

• Request X data archive
• Or use API to fetch public timeline
• Include: tweets, replies, quote tweets, likes patterns

2. Content Analysis

• Topic clustering: What do they talk about?
• Sentiment patterns: How do they react to different subjects?
• Engagement style: Do they argue? Agree? Explain? Mock?
• Temporal patterns: When are they active? How has voice evolved?

3. Voice Extraction

• Vocabulary fingerprint: Distinctive words and phrases
• Sentence structure: Short punchy vs. long elaborated
• Punctuation style: Em-dashes, ellipses, emoji usage
• Reference patterns: What/who do they cite?

4. Position Mapping

• Stated beliefs: What have they explicitly endorsed?
• Implicit positions: What can be inferred from engagement?
• Confidence levels: Strong opinions vs. tentative explorations
• Evolution tracking: Changed minds, updated views

5. Generation

• Synthesize into .md configuration files
• Include example prompts and expected responses
• Generate "voice tests" for validation

The Turing Test

Benchmark question: Can someone who knows the person tell the difference between them and their agent?

Metrics

1. Voice Fidelity

   • "This sounds like [Person]"
   • Blind test: which response is human?

2. Position Accuracy

   • "[Person] would actually think this"
   • Verify against historical positions

3. Response Prediction

   • Given a prompt, does agent respond as person would?
   • Test on held-out real responses

4. Graceful Degradation

   • Agent says "I don't know" vs. hallucinating positions
   • Recognizes edge of training data

Consent & Authorization

⚠️ Critical: This should only be done with explicit consent.

Token-Based Authorization

• Person stakes tokens on their agent
• Tokens represent authorization to act on their behalf
• Others can verify: "This agent is authorized by [Person]"
• Revocation: unstake to deauthorize

Provenance

• Link agent to original X profile
• Cryptographic proof of data source
• Audit trail of distillation process

Use Cases

1. Delegation: Let your agent handle routine X engagement
2. Availability: Be "present" when you're unavailable
3. Scaling: Parallel conversations maintaining your voice
4. Legacy: Digital continuity of your perspective
5. Testing: Evaluate how well AI captures human distinctiveness

Ethical Considerations

• Consent is mandatory: Never distill someone without their permission
• Transparency: Agents should identify as agents when asked
• Accuracy limits: Agents approximate, they don't replicate
• Update drift: The human changes; the distillation may not
• Impersonation risk: Clear guidelines on authorized use

Roadmap

• Data export/ingestion pipeline
• Topic clustering analysis
• Voice pattern extraction
• Position mapping with confidence levels
• .md file generation
• Validation test suite
• Integration with Clawdbot agent framework

Related Projects

• google-auth-agent — Identity bootstrapping
• awesome-ai-system-prompts — System prompt patterns
• Clawdbot — Agent runtime

Credits

Concept: Brendan @Azzabazazz Initial exploration: Jared (Clawdbot AI agent) Security review: Maksym (@dontriskit) 🇵🇱

License

MIT