AI Guides › Claude Mastery
The Habit Of Asking "What Would Change Your Answer?"
By Nigel Guy · 2 min read
Most people test a Claude answer by re-reading it more carefully, looking for
a wobble in the tone or a sentence that doesn't sit right. That's a weak test.
Confident, well-structured prose is not the same thing as a well-supported
conclusion, and re-reading the same words a second time rarely reveals which
one you're holding.
The rule: instead of asking whether an answer is right, ask what would
change it — a recommendation that can't name the fact that would flip it is
closer to a guess with good formatting than to a considered view.
The mechanism: the flip check
- Get the answer first, unedited. Don't hint at doubt in your original
question — you want the answer Claude would give anyone, not one already
defending itself.
- Ask directly: "what single piece of information, if it turned out to be
false, would change this recommendation?" A good answer names something
specific — a number, a date, a fact about your situation. A bad answer
restates the original reasoning in different words, which is the tell that
there wasn't much load-bearing evidence underneath it.
- Check whether that thing is actually true. This is the point of the
exercise — you've just been handed the one fact worth verifying, instead
of vaguely worrying about all of them.
- If nothing named would change it, treat that as information too. An
answer with no flip condition is either genuinely robust to the details,
or it's pattern-matching rather than reasoning about your specific case —
and you can usually tell which by whether the named "unshakeable" reason
is generic or specific to what you actually told it.
- For a genuinely important decision, run it twice. Ask what would
change the answer, then ask what would change that. Two layers is
usually enough to hit either bedrock or air.
What to skip
Skip this for anything low-stakes — a rewritten email, a first-draft
paragraph, a recipe substitution. The flip check earns its two minutes on
decisions you'd actually regret getting wrong: a purchase, a plan you're
about to commit time to, advice you're about to pass on to someone else.
Don't bother running it as a ritual on every single reply; it's a tool for
moments of doubt, not a tax on every conversation.
Guardrails
- A named flip condition is a sign of a more legible answer, not a proof of
a correct one. Claude can still name a plausible-sounding but wrong
condition with the same fluency as a right one.
- This works best on factual or analytical questions with a checkable
condition. It's much weaker on genuinely subjective calls ("which design do
you prefer") where there may be no single fact that would flip things.
- Verifying the named fact is still your job. Asking the question surfaces
what to check; it doesn't check it for you.
All 751 AI guides · JulieMango plans from £17/mo