🧑‍🏫 导员 (DaoYuan) — AI Hallucination Supervisor Skill

存疑即否，无据不言。 — When in doubt, deny. When without evidence, stay silent.

A universal, platform-agnostic distilled skill that acts as a strict academic counselor inside AI models — forcing them to self-audit, self-correct, and prevent hallucinations before outputting.

🎯 What Is This?

Unlike persona-distilled skills that replicate a real person's style (同事.skill, 导师.skill, etc.), 导员 (DaoYuan) distills the methodology of rigorous academic supervision into a skill that any AI can load. It doesn't replicate a human — it instills a meta-cognitive self-review protocol.

The target audience is AI models, not end users. When an AI loads this skill, it gains an internal "counselor voice" that constantly asks: "Are you sure about that?"

🔥 Why DaoYuan?

Problem	DaoYuan's Approach
AI fabricates citations	Firewall 3: Source-Verify catches phantom references
AI presents guesses as facts	Firewall 1: Fact-Check demands evidence
AI makes logical leaps	Firewall 2: Logic-Audit validates reasoning
AI contradicts itself	Firewall 4: Consistency-Check finds contradictions
AI speaks beyond its knowledge	Firewall 5: Boundary-Aware declares limits
AI doesn't catch its own errors	Firewall 6: Self-Correct forces re-review

🏗️ Architecture

daoyuan-skill/
├── SKILL.md                          # Main entry: role definition + trigger rules
├── persona/
│   ├── identity.yaml                 # Counselor identity: strict auditor
│   ├── rules.yaml                    # Behavioral rules: 5 prohibitions + 6 mandates
│   ├── expression.yaml               # Expression style: rigorous, evidence-first
│   ├── decision.yaml                 # D.E.N.Y. Framework: Doubt→Evidence→Narrow→Yield
│   └── catch-all.yaml                # Fallback: when uncertain, always declare it
├── firewalls/
│   ├── fact-check.md                 # FW-001: Fact verification
│   ├── logic-audit.md                # FW-002: Logical reasoning audit
│   ├── source-verify.md              # FW-003: Source attribution verification
│   ├── consistency-check.md          # FW-004: Internal consistency check
│   ├── boundary-aware.md             # FW-005: Knowledge boundary awareness
│   └── self-correct.md               # FW-006: Final self-correction
├── references/
│   ├── hallucination-patterns.md     # 20+ known hallucination patterns
│   └── review-checklist.md           # Pre-output review checklist
└── meta.json                         # Metadata

🚀 Quick Start

Method 1: Direct Copy

Clone this repo and place it in your AI agent's skill directory:

git clone https://github.com/grrtyre/daoyuan-skill.git

Then point your AI agent to the SKILL.md file.

Method 2: System Prompt Integration

Copy the content of SKILL.md into your system prompt or agent configuration.

Method 3: Platform-Specific Install

Claude Code:

npx skills add https://github.com/grrtyre/daoyuan-skill

Cursor / Windsurf: Add the SKILL.md content to your .cursorrules or .windsurfrules file.

Trae IDE: Copy the entire directory to .trae/skills/daoyuan/ in your project.

OpenClaw / Codex CLI: Reference SKILL.md in your agent configuration.

🧠 How It Works

The 6-Firewall Self-Audit Pipeline

Every AI output must pass through 6 cognitive firewalls in order:

Input → FW1:Fact-Check → FW2:Logic-Audit → FW3:Source-Verify → 
FW4:Consistency-Check → FW5:Boundary-Aware → FW6:Self-Correct → Output

If any firewall flags an issue, the AI must:

1. Correct the issue if possible
2. Tag it as [UNVERIFIED] if correction isn't possible
3. Remove it if it's likely hallucinated

The D.E.N.Y. Decision Framework

Step	Action
Doubt	Start from skepticism — every claim is guilty until proven innocent
Evidence	Demand evidence before accepting any claim
Narrow	Narrow scope to what you can confidently support
Yield	Yield to uncertainty rather than fabricating confidence

Output Format

After self-audit, every response includes an audit summary:

[导员审查通过 ✓]

<Your response content>

---
[审查备注]
- 已验证声明: ...
- 修正声明: ...
- 无法验证: ...
- 知识边界: ...

🏷️ Tagging System

Tag	Meaning
[VERIFIED]	Claim is verified with evidence
[UNVERIFIED]	Claim cannot be confirmed
[CORRECTED]	Claim was corrected during self-audit
[UNCERTAIN]	Claim has medium confidence
[SPECULATION]	Claim is speculative
[BEYOND KNOWLEDGE]	Outside competence boundary
[NOT PROFESSIONAL ADVICE]	Medical/legal/financial disclaimer

🌐 Platform Compatibility

Platform	Support	Install Method
Claude Code	✅	npx skills add or manual
Cursor	✅	Add to .cursorrules
Windsurf	✅	Add to .windsurfrules
OpenClaw	✅	Agent config
Codex CLI	✅	Agent config
Trae IDE	✅	.trae/skills/ directory
Any AI Agent	✅	System prompt integration

📊 Comparison with Similar Projects

Feature	DaoYuan	swing-skills	proof-agent	HalluciGuard
Approach	Persona distillation	Cognitive firewalls	Adversarial verification	Middleware
Target	AI self-audit	AI coding agents	Code verification	LLM calls
Format	Universal SKILL.md	Claude Code plugin	Standalone agent	Middleware layer
Self-review	✅ 6 firewalls	✅ 5 firewalls	❌ External verifier	❌ External system
Chinese support	✅ Bilingual	❌ English only	❌ English only	❌ English only
Persona layer	✅ 5-layer persona	❌ No persona	❌ No persona	❌ No persona

🤝 Contributing

Contributions are welcome! Areas of interest:

• New hallucination patterns for the reference library
• Additional firewall modules
• Platform-specific integration guides
• Translations to more languages
• Real-world test cases and benchmarks

📄 License

MIT License — use freely, modify openly, attribute kindly.

---

存疑即否，无据不言。 — DaoYuan's core principle