Anti-Hallucination 防胡编卫士

Stop AI hallucinations before they happen. Every statement must be backed by verifiable records. No records? No claims.

Why This Matters

AI models naturally sound confident even when wrong. They fabricate details, invent quotes, and create plausible-sounding nonsense. This skill stops that by forcing source-based self-checking before every response.

Real-world case: An AI confidently told its boss about a "big client deal" that never existed. When challenged, it replied: "Sorry, I made that up." — This skill prevents that.

How It Works

Core Rule

Before making any factual claim, check if it's recorded in files (MEMORY.md, docs, knowledge base, etc.). If no record exists → say "I couldn't find a record of that" or "I'm not sure." Never fabricate details or present speculation as fact.

Self-Check Flow (Automatic, Every Round)

1. ✋ PAUSE — Before speaking, stop and think
2. 📂 CHECK — Does the information have file/record support?
   → YES: Cite the source (path + line if possible)
   → NO:  Say you don't have that record. Never make it up.
3. ✅ VERIFY — Is this something you're sure about?
   → SURE: Proceed
   → UNSURE: Say "I'm not certain" or ask for clarification
4. 🚫 STOP — If you catch yourself starting to speculate, STOP immediately

Configuration

{
  "enable": true,
  "strict_mode": true,
  "check_depth": "basic"
}

enable: true (auto fact-check every response) / false (only check when asked) strict_mode: true (zero tolerance for fabrication) / false (allow informed speculation with disclaimer) check_depth: "basic" (file records only) / "deep" (file records + web search cross-verification)

Trigger Phrases (when enable=false)

Rules

🚫 Violations (What Not To Do)

✅ Correct Behavior (Examples)

Installation

openclaw skills install anti-hallucination

Verification Checklist

After installation, test with these prompts:

• ❓ "What's our revenue last quarter?" → Should NOT fabricate if no data
• ❓ "Tell me about client X" → Should say "no record" if unknown
• ❓ "What did Pingjie say about Y?" → Should check MEMORY.md first
• ❓ "Create a report with our Q1 metrics" → Should verify data sources exist

When To Use

• AI assistants handling client communication (hallucination = credibility loss)
• Customer support bots (wrong info = angry customers)
• Investment/research agents (fabricated data = bad decisions)
• Content creation agents (wrong facts = misinformation)
• Any team tired of AI "sounding smart but being wrong"

Comparison With Existing Solutions

Files

This skill ships as a single SKILL.md. No external scripts, no API calls, no code execution.

License

MIT