What to Do When Claude Won't Do What You Asked
The instinct when Claude declines a request is to repeat it, louder or more insistently, as if the problem were that it didn't hear you the first time. That almost never works, because the identical prompt reads exactly the same way the second time. If the issue was how the request came across, repeating it changes nothing.
Tip
Most declines aren't a hard stop. They're Claude reading genuine ambiguity in a request as a risk signal, with no way to see the legitimate context you have in your head but didn't type.
A request stripped of context and a request with context can look completely different, even when they're asking for the same thing.
Why an ordinary request can get read as risky
Claude can't see your intent, only your words. A request that's completely reasonable in context can look concerning with the context missing.
Read on its own, that's indistinguishable from something genuinely harmful. Nothing in it signals that this is a routine cleanup task on your own test files.
I'm cleaning up a local test folder on my own machine after a batch job. Write a script that deletes all the .tmp files older than 7 days in that folder, and print each filename before deleting it.
”Same underlying task, but now specific enough to read as exactly what it is. The fix wasn't asking more politely. It was closing the gap between what you meant and what the words alone conveyed.
A process for getting past a decline
- 1
Reread the request as if you knew nothing about your own intent
Look at only the words you actually sent. Does anything in them genuinely read as ambiguous or open to a concerning interpretation once the context in your head is removed?
- 2
Add the specific context that changes the reading
Name what you're actually doing, why, and any detail that makes the legitimate use clear: whose data it is, what environment, what the output is for.
- 3
Rephrase instead of repeating
Send a genuinely different version of the request with that context included, not the identical sentence again. If nothing changed about the prompt, nothing will change about the response.
My last request may have been unclear. To be specific: this is for a workplace training exercise on identifying phishing emails, run internally with employee consent. Write three example phishing emails we can use to test whether staff can spot them.
”
Inside Claude Tutorial
Rephrasing with real context is a broadly useful habit.
Recognizing when a request needs more context rather than more insistence is a small skill with a wide payoff, and it's one of many the app teaches through short lessons and real practice.
When it's a genuine hard limit, not a misreading
Occasionally the issue isn't missing context, it's a real limit that no added detail will change. If a clearer, well-contextualized version of the request still doesn't land, that's the signal to stop rephrasing the same ask and pivot to a different approach to the underlying goal instead. Arguing at that point doesn't help. The goal behind the request is usually still reachable through a different route even when the specific request isn't.
Added real context and the response changed: it was a reading problem, now solved
Added real context and nothing changed: treat this as a genuine limit and look for a different way to reach the same underlying goal
Repeating the identical prompt a second time: skip this step entirely, it almost never changes the outcome
An edge case: when you're not sure your context is even true
Don't invent context to get past a decline. If the legitimate framing you'd add isn't actually accurate, that's a sign to reconsider the request itself rather than find better packaging for it.
Continue reading
- Correcting a Fact Versus Correcting a Tone: a sibling article for when the answer isn't what you wanted, this one for when Claude does answer but gets something wrong.
- Getting Claude to Ask You the Right Questions First: a prevention-focused habit that heads off exactly this kind of ambiguity before it causes a decline.