LLMs state wrong facts with total confidence. Chain-of-Verification (CoVe) is a prompting pattern that makes the model catch its own mistakes — by drafting, then fact-checking each claim independently, then revising. Here's a live walkthrough.
🔬 Watch it catch a wrong fact: https://dev48v.infy.uk/prompt/day15-chain-of-verification.html
Four stages
- Draft — the model answers normally (e.g. "5 politicians born in NYC"). One entry is subtly wrong.
- Plan — it generates one verification question per claim ("Was X born in NYC?").
- Verify independently — it answers each question in a fresh context. This independence is the whole trick: if it re-checked while staring at its draft, it'd just repeat the error ("snowballing").
- Revise — it regenerates the answer, dropping/fixing whatever failed verification.
Why it works
Splitting one big claim into small, isolated checks turns "sound confident" into "answer this single fact." Small factual questions are much easier to get right, so the revision is grounded in verified pieces.
The cost
Several LLM calls per answer — worth it for fact- and list-heavy responses, overkill for casual chat.
🔨 Full pattern (draft → plan questions → answer independently → revise) on the page: https://dev48v.infy.uk/prompt/day15-chain-of-verification.html
Part of PromptFromZero. 🌐 https://dev48v.infy.uk
Top comments (0)