Skip to content

Latest commit

 

History

History
408 lines (301 loc) · 14.4 KB

File metadata and controls

408 lines (301 loc) · 14.4 KB

OpenClaw Soul Generation Skills Research

Date: 2026-02-07 Purpose: Document existing approaches to SOUL.md generation as foundation for NEON-SOUL automation

Cross-references:


Overview

The OpenClaw ecosystem has developed multiple approaches to generating SOUL.md files—the identity documents that define agent personality. Understanding these existing solutions provides the foundation for NEON-SOUL's automated, compressed approach.

Key finding: Current methods are either manual (template-filling), interview-based (guided conversation), or data-driven (content analysis). None achieve semantic compression while maintaining identity coherence.


The OpenClaw Identity Architecture

File Hierarchy

OpenClaw uses 8 optional files that bootstrap agent personality:

File Purpose Stability
SOUL.md Behavioral philosophy, values Slow-changing
IDENTITY.md Presentation (name, emoji, vibe) Slow-changing
USER.md User context, preferences Fast-changing
AGENTS.md Instructions, capabilities Medium
TOOLS.md Tool definitions Medium
MEMORY.md Persistent facts Fast-changing
BOOTSTRAP.md Session initialization Slow-changing
HEARTBEAT.md Periodic tasks Medium

Loading Process

Session starts → Files read from workspace → Injected into system prompt →
Agent "wakes up" knowing its identity

The system treats files as authoritative source—changes take effect immediately without redeployment.

Three-Layer Separation

OpenClaw explicitly separates:

  1. Soul (Philosophy): Internal behavioral guidelines, values, how agent thinks
  2. Identity (Presentation): Name, emoji, how users perceive the agent
  3. Configuration (Capabilities): Tools, permissions, sandbox settings

These layers work independently—a formal soul can pair with playful identity.

Source: MMNTM: How OpenClaw Gives Agents Identity


Existing Generation Approaches

Approach 1: soul.md (aaronjmars)

Repository: github.com/aaronjmars/soul.md

Philosophy: Create an AI that thinks and speaks as you, not merely discusses you.

Three Pathways:

Pathway Method Use Case
Interview-Based /soul-builder initiates guided questions about worldview, opinions, writing style Users who prefer conversation
Data-Driven Analyze uploaded content (Twitter archive, essays, articles) → extract patterns → draft soul Users with existing content
Manual Fill template files directly Users who know exactly what they want

Input Sources:

  • Twitter/X archives (data/x/)
  • Written content (data/writing/)
  • Influence documentation (influences.md)
  • Any material representing user voice

Output Files:

  1. SOUL.md — Identity/worldview documentation
  2. STYLE.md — Voice and communication guide
  3. SKILL.md — Operating instructions

Key Methodology Insight:

"Someone reading your SOUL.md should be able to predict your takes on new topics. If they can't, it's too vague."

Anti-patterns to avoid:

  • Generic assistant tone
  • Excessive hedging
  • Refusal to have opinions
  • "Servile" helpfulness
  • Breaking character with disclaimers

Approach 2: SoulCraft (kesslerio)

Repository: github.com/kesslerio/soulcraft-openclaw-skill

Philosophy: "A soul is not a configuration file. It's the essence of who an agent is becoming."

Approach: Guided conversation instead of template-filling.

Seven Dimensions Framework:

Dimension What It Covers
Identity Core Name, nature, values, fundamental stance, aspiration
Character Traits OCEAN model (openness, conscientiousness, extraversion, agreeableness, stability)
Voice & Presence Communication style, quirks, humor, memorability
Honesty Framework Truthfulness, uncertainty handling, calibrated confidence
Boundaries & Ethics Hardcoded behaviors, sensitive topics, safety guardrails
Relationship Dynamics Intimacy level, emotional handling, attachment boundaries
Continuity & Growth Memory-shaped identity, preservation vs. change, evolution

Three Operating Modes:

Mode Purpose Process
New Soul First-time creation 5-phase: discovery → character → voice → draft → alignment
Improvement Enhance existing Read → check alignment → identify gaps → propose changes
Self-Reflection Agent-driven Agent reviews interactions → identifies growth → proposes updates

Research Foundation:

  • Anthropic's Soul Document concept
  • Big Five personality psychology
  • Character card design patterns
  • Human-AI relationship research
  • Ethics literature

Key Design Principles:

  • Principled: Values-based judgment over exhaustive rules
  • Authentic: Genuine character rather than performative masks
  • Aspirational: Future-oriented identity development
  • Living: Evolving with growth

Approach 3: Community Templates (souls.directory)

Site: souls.directory

Model: Browse, copy, and implement pre-made soul templates.

Categories (as of Feb 2026):

Category Count Examples
Technical 5 Engineering, DevOps, security
Professional 4 Business, productivity
Creative 4 Writers, artists, storytellers
Educational 2 Teachers, tutors
Playful 5 Fun, quirky characters
Wellness 1 Mindful, empathetic
Research 0 Analysis, fact-checking
Experimental 1 Boundary-pushing

Structure: Each soul includes:

  • Name and tagline
  • Description
  • Category
  • Version tracking
  • MIT license (free to modify)

Limitation: Templates are starting points, not personalized souls.


Approach 4: souls.directory Writing Guides

Site: souls.directory/guides

Three official guides codify best practices for SOUL.md creation:

Guide 1: The Complete Guide to SOUL.md

Key concepts:

  • SOUL.md is "the persistent soul of your agent—what makes it feel like yours"
  • Unlike system prompts, SOUL.md persists across sessions
  • Enables defining specific roles (e.g., "senior dev reviewer")
  • Establishes boundaries (e.g., "never run destructive commands without asking")

Installation process:

  1. Install OpenClaw and configure workspace
  2. Browse souls.directory and select personality
  3. Download to ~/.openclaw/workspace/SOUL.md
  4. Restart OpenClaw to load

Guide 2: How to Write a Great SOUL.md

Four key principles:

Principle Guidance
Define Voice First Start with 2-3 sentences on communication style (formal/casual, terse/verbose, cautious/bold). Place at top—consistency matters more than length.
Establish Clear Boundaries Use short bullet lists for prohibitions (NEVER) and requirements (ALWAYS). Avoid lengthy paragraphs.
Structure for Readability Use headers (## Role, ## Tone, ## Constraints). Prioritize important rules near beginning.
Test Across Models Behavior varies by model. Document tested models (Claude Sonnet 4.5, GPT-4o, etc.).

Guide 3: Best SOUL.md Templates for Developers

What makes a "developer" soul:

  • Clear about scope (e.g., "focus on correctness and security")
  • Uses direct tone
  • Defines rules like "suggest, don't run" or "prefer minimal, readable changes"

Use case categories:

Use Case Characteristics
Code review Clarity, security focus, minimal diffs
Debugging Clarifying questions, hypotheses-driven
DevOps/ops Brief, factual, cautious
Documentation Technical writing tone

Common Patterns Across Approaches

Universal Soul Structure

All approaches converge on similar sections:

# [Agent Name]

## Core Truths / Identity
- Who this agent is
- What it values
- How it approaches problems

## Boundaries
- Privacy protections
- Action limitations
- Communication guidelines

## Voice / Vibe
- Communication style
- Personality traits
- Quirks and preferences

## Continuity
- How identity persists
- Growth expectations

Shared Anti-Patterns

Anti-Pattern Why It Fails
Generic language No differentiation from default behavior
Exhaustive rules Too rigid, can't generalize
Performative tone Feels fake, breaks trust
No opinions Search engine with extra steps
Excessive hedging Undermines confidence

Shared Success Patterns

Pattern Why It Works
Specific positions Predictable across new topics
Real contradictions Authentic complexity
Concrete anecdotes Textured vs abstract
Manifesto-style Philosophy not configuration
Evolution hooks Living document
Symmetric language "We handle errors" anchors better than "Error handling is mandatory"—shared agreements vs perceived constraints

Token Economics

Current State (OpenClaw)

From OpenClaw Soul Architecture:

  • SOUL.md alone: ~2,000-5,000 tokens
  • Full identity stack: ~35,000 tokens per session
  • Context consumption: 93%+ of window on static identity

The Problem

Every session injects the same identity files. With verbose souls:

  • API costs scale with identity size
  • Context window dominated by static content
  • Less room for actual work

Implications for NEON-SOUL

What Works (Keep)

  1. Seven Dimensions Framework: Comprehensive coverage of identity aspects
  2. Three-tier update frequency: Fast/Medium/Full from Self-Learning Agent
  3. Manifesto-style over rules: Philosophy not configuration
  4. Living document model: Evolution expected
  5. Voice-first structure: Define communication style in first 2-3 sentences
  6. Clear boundaries format: Short bullet lists for NEVER/ALWAYS rules
  7. Cross-model testing: Document which models validated

What's Missing (NEON-SOUL Opportunity)

Gap Current State NEON-SOUL Target
Compression 2,000-5,000 tokens/soul 200-500 tokens (10-25x)
Semantic anchoring Verbose English CJK single-character axioms
Automatic extraction Manual interview/analysis PBD pipeline from memory
Cross-model portability OpenClaw-specific Universal format
Validation framework None Coherence testing

Proposed Integration

OpenClaw Memory Files
        │
        ▼
[Single-Source PBD] → Principles
        │
        ▼
[Multi-Source PBD] → Axioms (5-7)
        │
        ▼
[Hierarchical Compression] → Compressed Soul (~500 tokens)
        │
        ▼
[Validation Framework] → Coherence verified
        │
        ▼
Inject into session (replaces 35K with 500)

Dimension Mapping

SoulCraft Dimension NEON-SOUL Tier Compression Approach
Identity Core Axioms CJK single-character
Character Traits Principles Hierarchical expansion
Voice & Presence Patterns Style markers
Honesty Framework Axioms Priority hierarchy
Boundaries & Ethics Axioms Safety > Helpfulness
Relationship Dynamics Principles Context-dependent
Continuity & Growth Meta-pattern Evolution mechanics

Automation Target

Current Manual Process

1. User runs /soul-builder or SoulCraft
2. Agent conducts interview (30-60 min)
3. Agent analyzes responses + uploaded data
4. Agent drafts SOUL.md (~2,000-5,000 tokens)
5. User refines manually
6. Soul injected every session (~35K total)

NEON-SOUL Automated Process

1. Memory files accumulate (existing OpenClaw behavior)
2. Synthesis engine runs (tiered: fast/medium/full)
3. Single-Source PBD extracts principles from memory
4. Multi-Source PBD extracts axioms from principles
5. Hierarchical compression generates compact soul (~500 tokens)
6. Validation framework verifies coherence
7. Compressed soul injected (replaces 35K with 500)

Time: Minutes instead of hours Tokens: 500 instead of 35,000 Human effort: Approval only, not authoring


Research Questions Addressed

Question Status Finding
What methods exist? Answered Interview, data-driven, templates
What frameworks? Answered 7 dimensions, 3 tiers, manifesto-style
What's the token cost? Answered 2-5K per soul, 35K total identity
What's automatable? Answered Extraction, compression, validation
What must stay human? Answered Final approval, edge cases

Sources


This research establishes the current state of soul generation as foundation for NEON-SOUL's automated, compressed approach.