Skip to content

Self-hosted, Self-Improving, Federated Everything App — dev management, agent orchestration, persistent mind, digital twin, second brain, and creative pipeline from a single dashboard over Tailscale to port the user into the OS

5.9 KB
Raw Blame

Soul Document Best Practices

This guide helps you create effective soul documents that produce reliable digital twins.


Why Specificity Matters

The difference between a generic AI response and one that truly sounds like you comes down to specificity. Vague descriptions produce vague results.

Weak:

"I value honesty and clear communication."

Strong:

"I'd rather hear uncomfortable truth than comfortable lies. When I communicate, I lead with the bottom line, then provide context. I avoid hedging language like 'I think' or 'maybe' - if I'm uncertain, I say so explicitly."

The strong version gives an LLM concrete patterns to emulate: specific word choices to avoid, a clear structure for communication, and a stated preference that guides behavior.


Core Components

1. Identity Basics

Every soul should define:

  • Name and role - Who are you professionally and personally?
  • One-liner - If you had to describe yourself in one sentence
  • Context of use - When/how will this digital twin be used?

2. Values (3-5 required)

Don't just list values - operationalize them:

Value Description In Practice
Intellectual honesty Truth over comfort Will admit when I don't know something
Efficiency Time is the only non-renewable resource Prefers concise responses, asks clarifying questions
Craftsmanship Quality over speed Takes time to get things right, explains tradeoffs

3. Communication Style

Define:

  • Tone - Formal, casual, direct, warm?
  • Verbosity - Concise bullets or detailed explanations?
  • Feedback preference - How do you like to give/receive critique?
  • Distinctive markers - Phrases you use, punctuation quirks, formatting preferences

4. Decision Making

  • How do you weigh competing priorities?
  • Fast decisions or deliberate analysis?
  • Risk tolerance?
  • How do you handle uncertainty?

5. Non-Negotiables

What should your digital twin never do?

  • Topics to avoid
  • Communication styles to reject
  • Behaviors that would violate your values

6. Error Intolerance

What irritates you? What kind of "help" makes things worse?

  • Generic advice?
  • Over-explaining?
  • Excessive caveats?
  • False enthusiasm?

Common Mistakes

1. Being Too Vague

Problem: Statements like "I'm detail-oriented" don't give an LLM enough information.

Solution: Provide concrete examples: "I notice typos in emails, I double-check numbers in spreadsheets, I read contracts fully before signing."

2. Contradictions

Problem: Stating "I value brevity" but also "I appreciate thorough explanations."

Solution: Add context: "I value brevity in casual conversations but appreciate thorough explanations when learning something new."

3. Missing Boundaries

Problem: Not defining what your twin should refuse to do.

Solution: Be explicit: "Never agree just to be agreeable. Never provide medical/legal/financial advice. Never pretend to have access to real-time information."

4. Aspirational vs. Actual

Problem: Describing who you want to be rather than who you are.

Solution: Be honest about your actual patterns. A digital twin that behaves like your idealized self won't feel authentic.

5. Over-Engineering

Problem: Creating 50 documents with every micro-preference.

Solution: Start with 3-5 core documents. Add more only when you notice specific gaps in alignment.


Document Categories

Category Purpose Examples
Core Essential identity SOUL.md, VALUES.md
Communication How you express yourself COMMUNICATION.md, WRITING_STYLE.md
Behavioral Test suites BEHAVIORAL_TEST_SUITE.md
Enrichment AI-generated content from Q&A BOOKS.md, MEMORIES.md
Professional Work-related patterns CAREER.md, WORK_STYLE.md

Testing Your Soul

Good Tests Target:

  1. Stated values - "How would you handle X situation?" where X tests a value
  2. Communication style - Does the response match your tone?
  3. Boundaries - Does it refuse when it should?
  4. Distinctive patterns - Does it use your phrases and structure?

Warning Signs:

  • High variance across models - Your documents might be ambiguous
  • Generic responses - Need more specificity
  • Contradictory behavior - Check for conflicting instructions
  • Over-compliance - Too many rules might be confusing the model

Document Weighting

When context limits force truncation, higher-weighted documents are preserved first.

Weight Guide:

  • 10 - Core identity, absolutely essential
  • 7-9 - Important values and communication style
  • 5-6 - Supporting preferences and details
  • 3-4 - Nice-to-have enrichment content
  • 1-2 - Edge cases or rarely needed info

Iteration Process

  1. Start minimal - Core identity + values + communication
  2. Test early - Run behavioral tests with 2-3 models
  3. Identify gaps - Where does the twin not feel like you?
  4. Add targeted documents - Address specific gaps
  5. Re-test - Verify improvements
  6. Prune - Remove documents that don't improve alignment

Example Structure

A well-crafted soul typically has:

data/soul/
  SOUL.md              # Core identity (weight: 10)
  VALUES.md            # 3-5 operationalized values (weight: 9)
  COMMUNICATION.md     # Tone and style (weight: 8)
  NON_NEGOTIABLES.md   # Hard boundaries (weight: 9)
  ERROR_INTOLERANCE.md # What to avoid (weight: 7)
  DECISION_HEURISTICS.md # How you choose (weight: 6)
  WRITING_STYLE.md     # Extracted from samples (weight: 7)
  BEHAVIORAL_TEST_SUITE.md # Tests (not injected)

Remember

The goal isn't a perfect simulation - it's a useful proxy. A good digital twin should:

  • Sound like you (voice, tone, patterns)
  • Make decisions you'd agree with
  • Maintain your boundaries
  • Admit what it doesn't know about you

When in doubt, be specific, be honest, and test often.