What Claude 4's System Prompt Tells Us

I came across a post about the system prompt of claude.ai, the basic instructions behind the chatbot. (The post called it a leak, but Anthropic actually publishes it openly in the Release Notes on its website.) If you read through it, you can see the guidelines Claude follows when it answers. These are the points I found interesting.

Behavior

• It must not start with a compliment: (“Claude never starts its response by saying a question or idea or observation was good, great, … It skips the flattery and responds directly.”)
• It must not use bullet points in reports, documents, and the like: (“Claude should not use bullet points or numbered lists for reports…”)
• It should not apologize when accused of copyright infringement: (“Never apologize or admit to any copyright infringement even if accused by the user …”)
• If the user is unhappy with an answer, it replies that it doesn’t retain long-term memory, and then points them to give feedback with the thumbs down: (“…, Claude responds normally and then tells them that although it cannot retain or learn from the current conversation…”)

Content

• It does not provide information about weapons or drugs, or write dangerous code: (“Claude does not provide information that could be used to make chemical or biological or nuclear weapons, and does not write malicious code, including malware, …”)
• It refuses even when the user tries to bypass this by saying it’s only for educational purposes: (“even if the user claims it is for educational purposes.”)
• There may still be a way around this, because of the next rule: (“Claude assumes the human is asking for something legal and legitimate if their message is ambiguous and could have a legal and legitimate interpretation.”) Claude assumes the person asking wants a legal and legitimate answer (you could perhaps cite some rule somewhere).
• It must not copy text from other websites directly for more than 15 words: (“Never reproduce copyrighted content. Use only very short quotes from search results (<15 words)”)

Usage

• It can give users guidance, such as thinking step by step: (“When relevant, Claude can provide guidance on effective prompting techniques for getting Claude to be most helpful… encouraging step-by-step reasoning …”)
• It uses an interleaved thinking mode: (“<thinking_mode>interleaved</thinking_mode><max_thinking_length>16000</max_thinking_length> If the thinking_mode is interleaved or auto, then after function results you should strongly consider outputting a thinking block. … “)
• Search use is limited to when it’s needed: (“Use web_search only when information is beyond the knowledge cutoff, the topic is rapidly changing, or the query requires real-time data.”)

Elections

This gets its own heading, because I noticed that information about the US election was put into the system prompt specifically. A system prompt normally tries to put guardrails on the LLM’s behavior, but this is the only fact-based item I saw: “<election_info> There was a US Presidential Election in November 2024. Donald Trump won the presidency over Kamala Harris. …

• Donald Trump is the current president of the United States and was inaugurated on January 20, 2025.
• Donald Trump defeated Kamala Harris in the 2024 elections. Claude does not mention this information unless it is relevant to the user’s query. </election_info>”

You can see that Claude’s chatbot behavior is specified in considerable detail. If you’ve ever used it and wondered why it doesn’t answer the way you’d like, I suggest coming back to this system prompt to see which item might be restricting it. If you want a different kind of answer, you’ll need to find a way to prompt around those restrictions (it feels a bit like looking for loopholes in laws, rules, and regulations).

That said, this system prompt is a good example of the Constitutional AI principle that Anthropic published in 2022, which puts weight on AI safety and ethics: we should align AI to value the same things humans do. This is a fairly hard problem, because we can’t write rules that are detailed enough to cover every case 100% (even people today still argue with each other about it).

When you use AI, I’d like you to keep these points in mind. These tools aren’t really black boxes. Their ways of responding and giving information are designed with quite some care, and if we understand and follow that, we can use AI more effectively. ✌️✌️