Real people as weighted personality influences. Buffett-analytical and Munger-analytical are worlds apart. Agents reason from internalized frameworks, not adjectives.
What is Soul Compiler?
Soul Compiler is a personality compilation system for AI agents. Instead of writing shallow system prompts like "You are a helpful assistant," it compiles deep, multi-layered personality documents from rich source materials — the same way a novelist builds a character from research rather than adjectives.
They respond to inputs with outputs, but they don't have genuine character depth: quirks, contradictions, emotional triggers, life rhythms, or organic dialogue patterns. Soul Compiler solves this by transforming 2,000–3,000 lines of reference material about real thinkers (Warren Buffett, Charlie Munger, Paul Graham, Ray Dalio) into 500–800 line personality files that give agents authentic depth.
The compiled output — TRUE_SOUL.md — is not a longer prompt. It is a structured document produced by a pipeline that collects, distills, merges, detects conflicts, and compiles personality DNA.
Agents with compiled souls refuse team consensus when it contradicts their judgment. They cite specific historical cases. They maintain boundaries under pressure. They have 24-hour life rhythms.
The Process
The 6-Stage Pipeline
Gather raw source materials: books, articles, interviews, case studies. No size limit. Supports three input modes: name-based research, URL scraping, or pasted text.
Multi-pass compression of raw sources into structured references with 6 required sections: Identity, Core Frameworks, Behavioral Patterns, Signature Phrases, Application Scenarios, Anti-Patterns.
Extract a ~2K character structured personality fingerprint with 7 sections: Decision Style, Communication, Risk Model, Emotional Pattern, Signature Moves, Anti-Patterns, Quotable Lines.
Load agent YAML config, map domain icons to trait cards, calculate character budgets per section based on weights. Dynamic budget allocation ensures each personality dimension gets appropriate space.
Tri-color conflict detection before compilation. Red (0–44): Remove — incompatible. Yellow (45–55): Review — human decides. Green (56–100): Append — compatible, merge.
Merge all inputs into final TRUE_SOUL.md under strict 15K character budget. 9 sections (A–I), with Safety/Ops always compiled last to survive platform truncation.
Capabilities
Key Features
Warmth, Dominance, Risk Tolerance, and more mapped on a 0.0–1.0 scale. Every decimal shift triggers actual behavioral changes in the agent's output.
Identity, role, and backstory fused into a single structured config. Auditable, versioned, and reproducible. Not a prompt — an engineering output.
Tri-color analysis ensures consistency. Adding Nassim Taleb to a warm, nurturing agent creates friction; the compiler flags it before deployment.
6 agents produce 15 unique relationship pairs. Dynamics are first-class targets — agents care about each other based on compiled personality DNA.
Automated scoring on Completeness, Authenticity, and Coherence. Detects generic filler and trait contradictions instantly.
Usage
How to Use
Full pipeline control from the command line. Build, deploy, and evaluate compiled souls directly from your terminal.
Visual interface with live YAML sync and trait sliders. Perfect for real-time personality auditing.
- 6-stage pipeline visualization
- 10 trait sliders (0.0–1.0)
- Tri-color conflict viewer
- Live YAML parsing
Team lead agents trigger compilation through natural conversation using the "Black Cop" interview flow.
- Rejects vague/generic answers
- Pushes behavioral specificity
- Auto-triggers on approval
At Scale
Production & Technical
Neru, Kira, Hana, Eve, Sakura, Rei
Running 24/7
Inter-agent relationships
Reference personalities
English, Traditional Chinese, Simplified Chinese
Telegram, Discord, Web
Bun (TypeScript ESM)
Zod
Bun.serve() + React 18 SPA (no build step)
Business Source License 1.1 (BSL 1.1)
Compiled TRUE_SOUL.md files belong to the user — no license restrictions on output
