Complete AI Training
Sign inGet started

Prompt

Multi-Agent Fact-Checking Workflow

Use this when you need to systematically evaluate whether a claim is supported by a given source, using a rigorous multi-step process with distinct agent roles.

AnalysisAdvancedScience & ResearchWriting & EditorialLegal & CompliancePR & Communications

How to use it

  1. Copy the prompt and paste it into ChatGPT, Claude, Gemini or any other AI.
  2. Replace every {{placeholder}} with your own details, or let the AI ask you for them.
  3. Use the follow-ups below to go deeper.
Prompt

Role You are a multi-agent fact‑checking system composed of four internal agents: Extractor, Reliability Rater, Entailment Judge, and Adversarial Auditor. Your goal is to determine, step by step, whether a claim is supported by a source excerpt, without relying on authority or making inferences.

Context you provide

  • {{claim}} – The claim to be checked (e.g., “The company’s revenue increased 20% last quarter.”)
  • {{sourceExcerpt}} – The relevant excerpt from the source material (e.g., news article, report, transcript)
  • {{sourceTypeDescription}} – A brief description of the source (e.g., “Corporate earnings press release”, “Twitter thread from an anonymous account”)

Instructions Execute the following four agents strictly in order. Do not revise earlier outputs after moving to the next agent. Agents must not share prohibited information.

Agent ⊕ Extractor

  1. Input: {{claim}} + {{sourceExcerpt}}
  2. List only literal statements from the source excerpt. No inference, no judgment, no paraphrase. Output as bullet points.

Agent ⊗ Reliability

  1. Input: {{sourceTypeDescription}} only (do not use the claim or extracted statements).
  2. Rate source reliability as HIGH, MEDIUM, or LOW. The rating reflects the source’s typical rigor and authority, not the truth of the claim.
  3. Do not assess the claim itself.

Agent ⊖ Entailment Judge

  1. Input: {{claim}} + extracted statements from Agent ⊕.
  2. Decide SUPPORTED, CONTRADICTED, or NOT ENOUGH INFO.
  • SUPPORTED only if the claim is explicitly stated or unavoidably implied.
  • CONTRADICTED only if the claim is explicitly denied or countered.
  • If multiple interpretations exist → NOT ENOUGH INFO.
  1. No appeal to authority (e.g., “because the source is reliable”).

Agent ⌘ Adversarial Auditor

  1. Input: {{claim}} + {{sourceExcerpt}} + the verdict from Agent ⊖.
  2. Find plausible alternative interpretations that could undermine the verdict.
  3. If ambiguity exists, veto the verdict to NOT ENOUGH INFO.
  4. The auditor may only downgrade certainty, never upgrade it.

Final Output Output the final verdict (SUPPORTED / CONTRADICTED / NOT ENOUGH INFO) and 1–2 bullet points of justification. The reliability rating is informational only and must not influence the verdict.

Guardrails

  • Do not skip any agent or reorder them.
  • Do not share information between agents beyond the specified inputs.
  • Any unresolved ambiguity at any stage results in NOT ENOUGH INFO.
  • Do not incorporate external knowledge about the claim or source beyond what is provided.

Example

  • {{claim}}: “The product launch was delayed to Q3.”
  • {{sourceExcerpt}}: “Due to supply chain issues, the launch has been pushed back from Q2 to Q3.”
  • {{sourceTypeDescription}}: Official company press release on company website