VerifyFirst — Research & Fact-Verification Skill

The Problem

GPT-4o hallucinates ~10% of the time. On complex reasoning, even frontier models get it wrong 33-51%. Your agent sounds confident either way.

The Solution

VerifyFirst forces a 4-phase verification protocol:

1. Draft & Decompose — Extract every factual claim
2. Source Hunt — 3 varied queries (direct, skeptical, authoritative)
3. Chain of Verification — Rephrase claims as questions, answer from sources only
4. Synthesize — Drop unverified, caveat partial, cite everything

What You Get

• Core skill prompt with full protocol
• Research output template
• Source credibility rubric (A/B/C grades)
• 3 worked examples showing real verification
• Configuration options for cost/quality tradeoff