Chain of Verification (CoVe) — Prompting Guide & Examples
Chain of Verification asks the AI to generate an answer, then systematically verify each claim in its response by asking itself targeted follow-up questions. This dramatically reduces hallucinations and factual errors.
How It Works
Four steps: (1) Generate initial response, (2) Extract verifiable claims, (3) Generate verification questions for each claim, (4) Answer those questions independently and revise the original response based on findings. Claims that fail verification get corrected or removed.
When to Use
Use CoVe for factual content generation, research summaries, technical documentation, and any task where accuracy matters. Critical for applications where hallucinations could cause harm — medical, legal, financial content.
Model-Specific Tips
ChatGPT / GPT-4
GPT-4 handles CoVe well. Structure the verification as explicit steps in your prompt. Use system prompts to enforce the verify-then-respond pattern.
Claude
Claude is excellent at CoVe due to its tendency toward honesty. Ask Claude to assign confidence levels and it will flag uncertain claims proactively.
Gemini
Gemini supports CoVe reasoning. Use structured prompts with clear verification steps. Gemini's grounding capabilities can enhance verification.
Pros & Cons
Pros
- ✓ Dramatically reduces hallucinations
- ✓ Self-verifying — catches factual errors
- ✓ Provides confidence levels for claims
- ✓ Essential for high-stakes content
Cons
- ✗ Very high token usage (4-5x baseline)
- ✗ Slower due to multi-step verification
- ✗ Model may verify against incorrect training data
- ✗ Can't verify truly novel information
Example Prompts
Write a summary of the key features of PostgreSQL 16. Now list every factual claim in your summary as a numbered list. For each claim, generate a verification question and answer it. Flag any claims you're uncertain about. Finally, produce a revised summary that only includes verified claims.
Explain the differences between React Server Components and traditional SSR. Verify each technical claim by questioning yourself: 'Is this actually true? What's my confidence level?' Revise any uncertain claims.
Describe the tax implications of converting a traditional IRA to a Roth IRA. List every financial/legal claim made. For each one, verify it against your training data. Mark confidence as HIGH/MEDIUM/LOW. Remove or caveat anything below HIGH confidence.