Chain-of-Verification (CoVe)
Reasoning
Level 2 · Advanced reasoningDraft an answer, generate verification questions about it, answer those questions, then revise the answer.
Chain-of-Verification takes an existing draft, generates a small set of specific verification questions targeted at catching an error in it, answers each independently, then revises the draft based on what those answers found. It's distinct from Self-Critique/Refine's more general review step: verification here happens through explicitly generated, checkable questions rather than a general 'review this' pass, which tends to catch specific factual errors a generic critique would skim past.
When to use it
Use it for factual or claim-heavy answers, where a specific wrong detail matters more than overall polish.
When to avoid it
Avoid it for subjective or creative tasks with no factual claims to verify — there's nothing for a verification question to check.
Advantages
- Catches specific factual errors a generic 'review this' pass tends to skim past
- Each verification question is independently checkable, not just a vague impression
Disadvantages
- Adds real length and turns compared to a single-pass answer
- Only as good as the verification questions generated — a shallow question set misses real errors
Examples
Draft: [a summary claiming a company was founded in 2015, has 500 employees, and is headquartered in Austin] Generate 3-5 specific verification questions that would catch an error in this draft. Answer each independently, then revise the draft based on the results.
Pairs well with
Try it
Fill in the placeholders below and see the compiled prompt update live.
The material to work from.
- 26
- Words
- 173
- Characters
- 0/1
- Placeholders filled
- 1 min
- Reading time
- Completeness0
- Clarity0
- Context0
- Specificity0
- Constraints0
- Format0
- Output0
- Constraints note0
- Success criteria0
- Scope focus100
- Verification step0
- Negative example0
- Confidence0
- Placeholder syntax100
- Markdown structure0
- Fill in the remaining placeholders.
- Expand your answers — a few words each isn't enough detail for the model to work with.
- Add background context so the model understands the situation.