AI Guides › Claude Mastery

The Two-Minute Audit Before You Trust A Claude-Written Number

By Nigel Guy · 2 min read

A number appears in Claude's answer — a total, a percentage, a projection, a count — formatted cleanly and stated plainly, and it's tempting to treat the formatting as a proxy for accuracy. A number that looks precise reads as if it must have been checked. It usually hasn't been, not in the way you'd check it yourself, and the two minutes it takes to verify the ones that matter is cheap against what an unchecked wrong one can cost.

The mechanism

The two-minute audit:

1. Ask where it came from. Was it calculated from data you supplied, pulled from something Claude already "knew," or estimated because nothing more specific was available? These are three very different levels of trustworthiness, and the answer to "where did this come from" often reveals which one you're looking at.
2. Redo the arithmetic yourself, roughly. Not the full calculation — a sanity check. Does the total look like it's in the right neighbourhood given the inputs? Order-of-magnitude checks catch a surprising share of real errors in under thirty seconds.
3. Check the number against one thing you already know. If a projected growth rate implies something wildly different from what you already understand about the underlying trend, that mismatch is worth more than a second read of the number itself.
4. Ask explicitly if it's an estimate, and if so, on what basis. Claude will generally tell you plainly when asked directly — the mistake is not asking, not that it hides the distinction when asked.
5. If it still matters, trace it to source. For anything that goes into a report, a client deliverable, or a decision with real consequences, follow the number back to the original data or document it supposedly came from, rather than trusting the intermediate calculation.

What to skip

Skip auditing every number in every casual conversation — a rough estimate for a low-stakes question doesn't need the same scrutiny as a figure going into a board deck. Match the audit to what the number is actually for. And skip assuming a confident tone means a checked number; tone and accuracy are unrelated, and treating them as linked is exactly the habit this guide is meant to break.

Guardrails

• This audit catches a meaningful share of errors quickly, but it isn't a substitute for a proper review process on anything financially or legally significant — those need a person with the authority and the source data to actually verify, not a two-minute gut check.
• A number can pass every step of this check and still be wrong if the underlying data it was calculated from was itself wrong or incomplete — the audit checks Claude's arithmetic and framing, not your source data's accuracy.
• Build the habit for the numbers that matter and let the low-stakes ones go — a two-minute check applied indiscriminately to everything just becomes friction, not rigour.

All 751 AI guides · JulieMango plans from £17/mo