Verification mistakes
Check summaries against the source and code by running it, ignore confident wording, and keep hard decisions for yourself.
-
How Do You Check a Claude Summary Against the Source? Have Claude tie each claim to a word-for-word quote, then look up the claims you'll rely on in the original. Level 2: Context
- 1Pull exact quotes
- 2Write from the quotes
- 3Match each claim to a quote
- 4Remove what has no quote
-
Which Decisions Shouldn't You Leave to Claude? Anything that's hard to reverse, affects other people, or depends on context Claude can't see. Level 2: Context
- Easy to undo, just youClaude drafts, you skim
- Affects other peopleClaude drafts, you decide
- Hard to reverseyou decide, and limit what tools can do
- High-risk areasa qualified person must review and can change it
-
How Do You Check Code Claude Wrote Before Shipping It? Run it, and ask for proof. Level 2: Context
- 1In the promptClaude decides when it's satisfied
- 2/goala separate model checks each turn
- 3Stop hookClaude Code enforces it
- 4CIchecks outside the session
-
Why Does Claude Sound Sure When It's Wrong? Claude writes in the same fluent tone whether a detail is right or wrong, so tone tells you nothing. Level 2: Context↑ cost if wrongFind a second sourcecostly, hard to checkAlways checkcostly, easy to checkTreat as a leadcheap, hard to checkQuick lookcheap, easy to checkhow easy to check →
Match the check to the claim
Different outputs fail in different ways, so one habit won't cover them. For a summary, read the quote against the claim. For a fact or number, open a source. For code, run it and read the output. For a decision, ask what Claude can't see. The cards in this lesson take these one at a time, each with a prompt that makes checking quicker.
Where checking can be automated
In Claude Code, part of this can be enforced by the tool. A Stop hook can keep Claude working until a check passes, and /code-review runs in a fresh subagent that didn't see your conversation. In the chat app, you do the checking, with help from citations when Claude uses web search. Neither replaces your judgment on work that matters.
How do I fact-check Claude's answers?
Ask for a source for each specific claim, and open the ones you'll rely on. Turn on web search for anything that changes. Anthropic notes that some links may not work, so confirm the page says what the answer says.
Are Claude's citations reliable?
They give you a place to check. Anthropic says its API citations are guaranteed to point to valid locations in the documents you provide, but you still need to read whether the passage supports the claim.
Should I test code that Claude writes?
Yes. Anthropic's Claude Code guide says Claude stops when work looks done, so give it a test, build or linter to run, and ask to see the command and its output.
Can Claude check its own work?
Partly. It can run tests and compare a draft to criteria, and Anthropic says this helps. Its Fable 5 guide says fresh-context verifiers tend to outperform self-critique, and high-stakes work still needs a qualified person.