AI Guides › Claude Mastery

Why Claude Sometimes Says No

By Nigel Guy · 2 min read

A refusal or a hedged, over-cautious answer tends to get read as a fault — the model being difficult, or overly nervous about something harmless. That reaction skips a genuinely useful step: a refusal is information about how the request looked from the other side, and that information is usually specific enough to act on, if you actually read it instead of just being annoyed by it.

The mechanism

1. Reread your own request as if you'd never seen the context in your head. A huge share of refusals come from a request that's genuinely ambiguous between an ordinary use and a concerning one, with all the disambiguating context living only in your own mind, not in the text you sent.
2. Add the missing context explicitly, once. If the request was ambiguous, state the legitimate purpose plainly — what it's for, who it's for, what you'll do with it — rather than assuming intent is obvious. This resolves a large share of refusals without changing what you're actually asking for.
3. Split compound requests apart. A request that bundles something ordinary with something borderline often gets treated as a whole. Separating them lets the ordinary part through cleanly and isolates exactly what's actually being questioned.
4. If it's still declined after that, take the decline as the answer for this framing, not a bug to route around. Repeating the identical request with more forceful wording is rarely what changes the outcome — a genuinely different, more specific framing sometimes will; sheer persistence generally won't.
5. For something you believe is a clear miss, say so plainly and move on to a different approach to the same underlying need, rather than spending the rest of the conversation arguing the point.

What to skip

Skip taking a refusal personally — it's a response to a request as written, not a comment on your character. And skip the instinct to disguise a request specifically to get past a decline; if the honestly-stated version wouldn't be answered, a disguised version getting through isn't a win, it's just a worse-documented version of the same outcome.

Guardrails

• Exactly where the lines sit changes as the underlying models and policies are updated, and reasonable people can disagree about where a specific line should be. This guide is about how to respond to a refusal, not an argument that every refusal is correctly placed.
• Some refusals are simply correct and shouldn't be engineered around at all — anything touching real harm to a real person is not a phrasing problem to solve.
• If a refusal seems clearly wrong for entirely ordinary, low-stakes work, that's worth reporting through the product's actual feedback channel, not just working around in the moment — it's the only way that specific miss gets fixed for the next person who hits it too.

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