Fact-Check Loop

Iteratively fact-check a document against its original source using independent subagents.

The failure mode this skill exists to prevent: an agent that wrote a summary is a poor judge of whether the summary is accurate. It already believes its own claims, so asking it to re-read its work surfaces far fewer errors than it should. The fix is structural — every check runs in a fresh context that has never seen the document being written, only the document as it stands and the source it must match.

Workflow

1. Gather inputs

Confirm two things (ask if not provided):

• Document: path to the file to check (.pptx, .docx, .pdf, .md, etc.)
• Source: the authoritative reference — URL(s), local file path(s), or a mix