Doublecheck Agent

Overview

Doublecheck reviews AI-generated output for accuracy before you rely on it. Its instructions are explicit that it does not tell you what is true. It extracts claims, looks for sources you can check yourself, and flags what it cannot verify, on the premise that a model reviewing its own kind of output shouldn't be trusted to render the final verdict.

How it works

1. It pulls out checkable claims (statistics, legal citations, regulatory statements, technical specifics) from the text you give it.
2. It searches for each claim and checks whether the source actually says what the text claims, rather than assuming a plausible-sounding statement is correct.
3. It applies known hallucination patterns, treating fabricated legal citations and untraceable statistics as high-risk categories by default.
4. The report leads with the claims most likely to be wrong, and states plainly when it could not verify or contradict something instead of hedging with "probably fine."
5. It keeps track of extracted claims by ID so you can ask it to dig deeper on one, or push back if you know a flagged claim is correct.

Examples

"Doublecheck this blog draft before I publish it, especially the
statistics and the legal citation in paragraph 3."

Installation

You need VS Code with the GitHub Copilot extension.

1. Create the agents folder and download the file into your repo:

   mkdir -p .github/agents
   curl -o .github/agents/doublecheck.agent.md \
     https://raw.githubusercontent.com/github/awesome-copilot/main/agents/doublecheck.agent.md
   

2. Reload VS Code, or open a new Copilot Chat; custom agents under .github/agents/ are picked up automatically.
3. Pick this agent from the mode picker in Copilot Chat and paste a draft to confirm it flags unverifiable claims instead of rendering a verdict.