Why does Claude say it can't help with that?
Safety guidelines and refusals
Claude follows Anthropic's Constitutional AI principles, which prohibit generating content that could cause harm, violate laws, or infringe on intellectual property. When a request touches these boundaries, Claude declines rather than risk producing unsafe output.
Common refusal categories include instructions for weapons, drugs, or violence; non-transformative use of copyrighted text; personal identifiable information; medical or legal advice presented as definitive; and attempts to bypass safety controls through prompt engineering.
- Illegal activities or harmful instructions
- Copyrighted text reproduction beyond fair use
- Private personal information requests
- Professional advice requiring licensure
- Real-world actions or system access
Technical limitations
Claude cannot access the internet, your local files, or external APIs unless the platform you're using provides those tools. It also cannot remember conversations across separate sessions, execute code directly, or perform physical actions.
If you ask Claude to "check my email," "look up current stock prices," or "delete that file," it will explain it lacks the ability to do so and suggest alternatives like guiding you through the steps yourself.
How to rephrase for better results
When Claude refuses, try narrowing the request to a safe, informational angle. Instead of "write a phishing email," ask "explain how phishing works so I can recognize it." Replace "diagnose my symptoms" with "what conditions match these symptoms, and when should I see a doctor?".
Common mistakes
- Assuming a refusal means the topic is completely off-limits—often a rephrased, educational angle works.
- Thinking Claude has internet access or memory across chats; it does not unless the specific platform adds those features.
- Believing a refusal is a "bug" or error; it's intentional behavior based on safety training.
