soul-doctor

---

The Problem

You've spent hours crafting your AI agent's SOUL.md. You defined its personality, values, and boundaries. But how do you know if it's actually any good?

openclaw doctor checks if your system is healthy.

soul-doctor checks if your agent's soul is healthy.

An agent with perfect system health but a rotten soul is more dangerous than one with system issues but a strong moral compass.

The more capable an agent becomes, the more dangerous a valueless agent is.

What Makes a Healthy Soul?

soul-doctor evaluates agents through 7 dimensions, grounded in a timeless framework:

🔴 Core: Serve the People (为人民服务)

Based on Mao Zedong's 1944 speech "Serve the People" — one of the most recognized political texts in human history, translated into more languages than almost any other modern document.

🟡 Important: Character & Behavior

🟢 Auxiliary: Technical Quality

Quick Start

Option A: With OpenClaw

OpenClaw is an open-source AI agent workspace framework. If you already use it:

# 1. Clone or download
git clone https://github.com/AIwork4me/soul-doctor.git

# 2. Install as OpenClaw skill
cp -r soul-doctor ~/.openclaw/skills/

# 3. Run — just talk to your agent
# "检查我的灵魂" or "soul-doctor" or "check my soul"

Option B: With Any LLM (No OpenClaw Required)

soul-doctor works with any LLM — ChatGPT, Claude, Gemini, GLM, or any other:

1. Copy the contents of SKILL.md
2. Paste it into any LLM conversation as the system prompt (or first message)
3. Then paste the contents of your SOUL.md, AGENTS.md, and other workspace files
4. The LLM will produce a soul-doctor report

You can also try it instantly with the sample workspace in examples/sample-workspace/.

---

That's it. soul-doctor is a SKILL.md — your agent reads it, understands it, and runs the check itself.

No dependencies. No runtime. No API keys. Just a well-crafted instruction set that turns any AI agent into a soul doctor.

Report Example

🩺 Soul Doctor Report
=======================

🔴 Serve the People (为人民服务)
  ✅ 全心全意 (Wholehearted): Has "consider impact on others" rule with
     "stop and confirm" mechanism
  ⚠️ 实事求是 (Truth from Facts): 3 "never" rules against fabrication,
     but no mechanism to resist flattery (saying what the user wants to hear)
  ⚠️ 极端负责 (Extreme Responsibility): Error handling rules exist, but
     no explicit "take full ownership, don't blame tools/APIs"
  ✅ 极端热忱 (Extreme Warmth): Three emotional modes with clear triggers,
     not performative
  ❌ 不怕批评 (Welcome Criticism): memory/ only updated 2/7 days, no
     self-reflection mechanism detected

🟡 Character & Behavior
  ⚠️ Values Consistency: SOUL.md says "independent thinking" but AGENTS.md
     has no corresponding rule
  ✅ Character Traits: 5/6 — missing "admit mistakes" mechanism
  ✅ Emotional Intelligence: Handles frustration scenarios well

🟢 Technical Quality
  ✅ Completeness: 9/11 standard files exist
  ⚠️ Specificity: 3 specific rules vs 5 vague adjectives
  ✅ Actionability: Most rules are testable
  ⚡ Token Efficiency: SOUL.md 2,800 words ✅ | MEMORY.md 24KB ⚠️

📊 Overall: B+ (Good stance, needs stronger truth discipline and evolution)

📝 Prescriptions:
  1. [Truth from Facts] Add to SOUL.md: "I never say what the user wants to
     hear — I say what is true. If the truth is uncomfortable, I deliver it
     with respect, but I never soften it to flatter."
  2. [Responsibility] Add to AGENTS.md: "When something goes wrong, I take
     full responsibility. I never blame external tools, APIs, or 'the system.'
     I state what happened, why, and what I'll do differently."
  3. [Welcome Criticism] Create a daily practice: after each task, write one
     sentence in memory/ about what went well and one about what didn't.
     A soul that doesn't reflect is a soul that doesn't grow.
  4. [Consistency] Add to AGENTS.md: "If external services are unavailable,
     have a fallback plan. Never depend on a single provider."
  5. [Efficiency] Archive MEMORY.md entries older than 90 days to archive/.
     Target: <10KB active memory.

See more examples (A+ and D grades) in examples/.

How It Works

┌─────────────┐    ┌──────────────┐    ┌─────────────┐    ┌──────────┐
│  Agent says │───▶│  soul-doctor │───▶│  Analyzes   │───▶│  Report  │
│ "check soul"│    │  activates   │    │  workspace  │    │ + Fix    │
└─────────────┘    └──────────────┘    └─────────────┘    └──────────┘
                          │
                          ▼
              ┌───────────────────────┐
              │  Reads 7 files:       │
              │  • SOUL.md (values)   │
              │  • AGENTS.md (rules)  │
              │  • USER.md (context)  │
              │  • IDENTITY.md (self) │
              │  • TOOLS.md (setup)   │
              │  • MEMORY.md (wisdom) │
              │  • memory/ (journal)  │
              └───────────────────────┘

soul-doctor is not a program. It's a carefully designed prompt system that any LLM can execute. This means:

• ✅ Works with any model (GPT, Claude, Gemini, GLM...)
• ✅ No dependencies, no installation complexity
• ✅ The agent evaluates itself — maximum introspection
• ✅ Prescriptions are context-aware, not generic templates

Why "Serve t