How to Use Claude for Fact-Checking: The Iris Skill Guide

There is an irony in using AI to check facts: the same tools that make misinformation cheap can also be confidently wrong themselves. So AI fact-checking only works if you use it the right way — not as an oracle that hands down verdicts, but as an investigator that interrogates a claim, separates what is established from what is asserted, and tells you honestly where the evidence runs out. Used like that, a tool like Claude, ChatGPT, or any AI chat becomes a genuine check on what you are about to repeat, rather than another source of confident nonsense.

Break the claim into checkable parts

“This statistic proves X” is usually several claims wearing one coat. The skill first decomposes it: what is the specific factual assertion, what is interpretation laid on top, and what would actually have to be true for the conclusion to hold. Most “false” claims are not lies — they are a true fact stretched into a wrong conclusion, and you only see that once it is taken apart.

Interrogate the source, not just the statement

A claim is only as good as where it came from. The skill helps you ask the investigator’s questions: who is saying this, what is their interest, is this a primary source or a summary of a summary, and when was it true? A confident claim with no traceable origin is a red flag, not a fact.

Separate “established” from “asserted”

This is the discipline that makes the whole thing safe. The skill marks each part of a claim by how well-supported it is — settled, contested, or unverifiable — rather than flattening everything into true or false. The honest answer is often “this part checks out, this part is disputed, and this part cannot be verified from here.” That nuance is the point.

Catch the rhetorical tricks

Plenty of misleading claims contain no false facts at all — they mislead through framing, cherry-picked dates, missing context, or a correlation dressed as cause. The skill is built to flag those moves, which the same investigative instinct in our market research guide applies to commercial claims.

Tell you what it cannot know

The most important output is the honest gap. The skill says plainly when it is reasoning rather than recalling, when something is past its knowledge, and when a claim simply needs a primary source it cannot reach — so you go and verify the load-bearing facts yourself. A fact-checker that admits its limits is the only kind worth trusting, and it pairs with the disciplined-sourcing habit in our research assistant guide.

Why a skill beats a one-off prompt

A loaded skill carries the method by default — decompose, interrogate the source, grade the evidence, flag the rhetoric, admit the gaps — so you get a real investigation every time rather than a quick confident answer. That discipline is the difference between checking a fact and just asking for one.

The honest limit

This bears repeating because it is the whole game: the skill can be wrong, and it cannot browse your specific source unless you bring it. Use it to structure the investigation and surface the questions — then confirm anything that matters against the primary source yourself. Used that way, AI fact-checking makes you harder to fool, including by the AI.