check AI claims against sources

The problem

AI-generated answers can sound reliable while containing unsupported details, source mismatches, outdated information, or claims that go further than the evidence actually allows. Manually checking every claim against its sources is slow and easy to miss.

What it does

• Extracts the important factual claims from an AI-generated answer, report, or summary.
• Splits compound claims into testable statements when needed.
• Maps each claim to the source material you provide.
• Labels claims as Supported, Partially Supported, Unsupported, or Unverified.
• Detects scope, date, version, and wording mismatches between a claim and its evidence.
• Highlights the highest-risk evidence gaps.
• Produces a safer corrected version that stays within what the available sources actually support.

Why use this instead of a general prompt?

A general “fact-check this” prompt can produce another confident summary without clearly showing what was verified. This skill uses a fixed claim-to-evidence workflow and requires every important claim to receive an explicit evidence status.

Typical output

• Claim-evidence audit table
• Supported / Partially Supported / Unsupported / Unverified status for each claim
• Up to three highest-risk evidence gaps
• A corrected version of the original AI answer

Use cases

• Checking AI-generated research or summaries against supplied sources.
• Verifying reports against data sheets, documentation, or source files.
• Reviewing AI-generated customer support or internal drafts before sending them.
• Comparing product claims against official manuals or specifications.
• Checking whether citations actually support the claims attached to them.

Known limitations

This skill can only verify sources the host agent can actually access and inspect. If a source is missing or inaccessible, the relevant claim is marked Unverified. It is a source-support audit, not a general truth detector.