How to use the Doublecheck verification skill for AI output

What is an agent skill, and what does Doublecheck make the agent do?

An agent skill is a folder containing a SKILL.md file that gives an AI agent specific instructions for a task. When the skill is installed and your request matches it, the agent follows those instructions.

Doublecheck runs a three-layer verification pipeline on AI-generated output. The first layer extracts claims and checks them for internal contradictions. The second layer searches for supporting or contradicting sources and gives you URLs to visit. The third layer runs an adversarial review for hallucination patterns and produces a structured report with source links for human review.

The skill does not decide what is true. Its stated goal is to extract every verifiable claim, find sources you can check yourself, and flag anything that looks like a hallucination.

• Claims are rated VERIFIED, PLAUSIBLE, DISPUTED, or FABRICATION RISK (the report template also uses UNVERIFIED).
• Inline mode adds a Verification section after a response; full mode produces a detailed report.
• For legal, regulatory, or compliance content, and for case or statute citations, the full report is always used.

How do you install the Doublecheck skill?

You can install the skill in either of two ways. Download the ZIP from this page and unzip it into one of these folders: ~/.claude/skills/ for your personal skills, or .claude/skills/ inside a project if you want the skill only for that project.

Alternatively, upload the skill in the Claude apps under Settings → Capabilities → Skills.

• Personal install: unzip into ~/.claude/skills/ so the doublecheck folder sits there.
• Project install: unzip into the .claude/skills/ folder of your project.
• App install: upload the skill under Settings → Capabilities → Skills.
• The folder should include SKILL.md and assets/verification-report-template.md.

What does the agent need to run Doublecheck?

The skill needs an agent runtime that supports SKILL.md skills and access to web search. The source files do not mention API keys or separate accounts.

The skill uses the network, because verification depends on web searches. It does not run scripts.

• Agent runtime that supports SKILL.md skills.
• Web search access.
• No API keys or separate accounts are mentioned in the files.

How do you trigger Doublecheck with natural requests?

Doublecheck has two modes. Active mode is persistent: invoke the skill without giving specific text, and the agent turns on verification for the rest of the conversation. It then adds a short verification summary after substantive responses. One-shot mode runs when you give the skill specific text or refer to earlier output and ask it to verify.

To get the detailed report on any response, say "full report". Similar phrases such as "verify that" or "doublecheck that" also work. To stop active mode, say "turn off doublecheck".

• "Doublecheck this summary and give me sources for every figure."
• "Verify that. Full report, please."
• "Doublecheck the citations in this memo: [paste text]."
• "Turn off doublecheck."

What do the verification results look like?

Inline results list each claim with its rating and a source link where one is found. Claims rated DISPUTED or FABRICATION RISK are called out at the top with a "Heads up" note, and the response moves to a full report.

Inline verification ends with a line suggesting you say "full report" for detailed sourcing. Full reports open with a summary table of counts by rating, then a flagged-items section with a recommendation for each problem claim, then all claims.

• VERIFIED: a source was found.
• PLAUSIBLE: no specific source found, but the claim seems reasonable.
• DISPUTED: a source contradicts the claim.
• FABRICATION RISK: the citation could not be found; check it before relying on it.

What are the safety and privacy considerations?

Doublecheck uses web searches, so the content and claims you check may be sent to search providers. Do not use it for offline or privacy-sensitive work where that is a problem.

Doublecheck verifies claims but cannot certify them, and search results may be incomplete or wrong. For legal or medical decisions, treat the output as a starting point for checking primary sources, not as a final answer.

The skill itself says verification does not apply to creative writing or code generation, and it will skip verification for those.

• Network use: yes. Web search is part of the process.
• Scripts: none are run.
• Do not rely on it for final legal or medical decisions.

Tips for getting useful results

Paste the exact text you want checked, or point to the specific response. The more precise the target, the easier it is to match claims to sources.

For drafts with case law or statutes, ask for the full report from the start rather than the inline summary. Open each source link yourself before you cite it.

• Name the parts you care about most, such as figures or citations.
• Check every source link independently, especially for FABRICATION RISK items.
• Turn active mode off when you only want normal responses.