chain-of-verification

chain-of-verification reduces hallucinations by explicitly breaking out the verification step: generate a response, extract individual claims, verify each claim independently, then revise with corrections.

When to use it

• Factual question answering where accuracy is critical
• Research summaries that will be cited
• Any task where the LLM might confidently hallucinate specific facts
• Legal, medical, or technical domains with high accuracy requirements

File structure

agents/my-cov/
├── AGENT.md    ← pattern: chain-of-verification, config
├── verify.md   ← extract verifiable claims
├── check.md    ← verify each claim
└── revise.md   ← revise with corrections

AGENT.md

---
name: fact-checker
description: Answers questions with verified facts.
version: "1.0.0"
pattern: chain-of-verification
call:
  model:
    role: thinker
---

Answer the question comprehensively. Include specific facts, dates, names, and statistics where relevant.

Call finish with key="done" and your answer as value.

verify.md

You have a generated response. Extract all verifiable factual claims.

For each claim, output it as a question that can be independently verified.
Format: one question per line, numbered.

Example: "The Eiffel Tower was built in 1889" → "Was the Eiffel Tower built in 1889?"

Call finish with key="done" and the numbered list of verification questions as value.

check.md

---
call:
  uses:
    - search-web
---

You have a list of factual claims to verify, expressed as questions.

For each question:
1. Search for authoritative sources
2. Mark the claim as VERIFIED, INCORRECT (with the correct fact), or UNVERIFIABLE

Format your response as a numbered list matching the input questions.

Call finish with key="done" and the verification results as value.

revise.md

You have the original response and a set of fact-check results.

If all claims are VERIFIED, return the original response unchanged.
Otherwise, revise the response to correct all INCORRECT claims.
Remove or qualify UNVERIFIABLE claims.

Call finish with key="done" and the fact-checked response as value.

Scaffold

tama add chain-of-verification my-agent

How it works

AGENT.md body → generate response with factual claims
      ↓
verify.md → extract claims as verifiable questions
      ↓
check.md → look up each claim (uses search-web if available)
      ↓
revise.md → correct errors, qualify unknowns