Ask the AI to check its own answer against a rule
Before the AI finishes, tell it what a right answer must satisfy and ask it to check its answer against that rule. Anthropic's prompting guide says this catches errors reliably, especially for math.
Why it matters
An AI can give a fast, confident answer that is wrong, for example a trip split that only counts one bill.
Anthropic's prompting docs say: append something like "Before you finish, verify your answer against [test criteria]." "This catches errors reliably, especially for coding and math."
The check works only if you name the rule, so say exactly what must be true.
Do this
- Paste the numbers or the text the answer depends on.
- Ask your question as usual.
- Add one line that names the rule: "Before you finish, check that [what must be true]."
- Check one result yourself before anyone sends money or acts on it.
A line to add at the end of your request
Before you finish, check your answer against this rule: [what must be true]. If it fails, fix it and show the corrected answer.
One thing to keep in mind: The check is only as good as the rule you give it. Name the rule, and check one number yourself.
The group-trip story in the video is an example; its numbers are part of the story. The docs describe Claude and add that Claude Opus 5 checks its own work well without being asked, so on that model you can leave the line out. Source: Anthropic: Prompting best practices, "Ask Claude to self-check".
Ask me anything about AI and cloudA paid call on your own setup: your agents, your bill, your plan.Book a call →