System prompt
In short
A system prompt is a set of hidden instructions given to an AI chatbot before you type anything. It tells the model who it is, how to behave and what it may or may not do, and it usually outranks what users ask.
- Also called
- System message, System instructions, Developer message
- Topic
- Using AI tools
- Level
- Beginner
- First described
- OpenAI's ChatML format, published on 1 March 2023 alongside the ChatGPT API, structured chats as a sequence of messages starting with a "system" message5.
Chat models read a conversation as a list of messages, each labelled with a role. OpenAI's Model Spec lists "system" messages added by OpenAI, "developer" messages from the app builder and "user" messages from the person typing, and says instructions with higher authority override lower ones1. OpenAI's API guide likewise says developer messages are prioritised ahead of user messages2.
Anthropic publishes the system prompts behind its Claude apps. It says they give Claude up-to-date information, such as the current date, at the start of every conversation, and encourage habits such as putting code in Markdown. Those prompts do not apply to the Claude API, where developers write their own3. Anthropic's docs add that setting a role in the system prompt, even a single sentence, focuses the model's behaviour and tone4.
A system prompt is not a security wall. It is just text the model was trained to weight highly. OpenAI tells its models to keep the verbatim text of system and developer messages private by default, but to share basic facts such as the model's identity and tools1. Determined users can still try to coax a prompt out, or to override it with a jailbreak or prompt injection.
Example
A UK bank's support chatbot might run on a system prompt like: "You are the assistant for Example Bank. Answer only questions about our accounts. Never give investment advice. If someone mentions fraud, direct them to our fraud team." Customers never see it, but it shapes every reply.
Why it matters
When two apps run on the same AI model but behave differently, the system prompt is usually why. It also explains why a chatbot may refuse topics that the underlying model would happily discuss elsewhere.
Questions people ask
Can I see a chatbot's system prompt?
Sometimes. Anthropic publishes the system prompts used in its Claude web and mobile apps3. Most companies keep theirs private, and OpenAI instructs its models not to reveal the verbatim text of system and developer messages by default1.
Is a system prompt the same as custom instructions?
They are related. Custom instructions are preferences you set in an app, which get added to the conversation for you. A system prompt is written by the company or developer and sits above your messages in priority1.
Can a system prompt be overridden?
It is designed not to be. OpenAI's Model Spec ranks system and developer instructions above user instructions1. In practice, jailbreaks and prompt injection attacks sometimes still get round them, which is why sensitive controls should not rely on the prompt alone.
Related terms
Tools that use it
- ClaudeAnthropic's chatbot, strong on writing and on pointing out holes in an argument, with a tight free allowance.
- ChatGPTOpenAI's general chatbot, the most used AI app in the UK and worldwide.
Sources
- OpenAI Model Spec (version dated 2026-08-18), sections on the chain of command and privileged information
- OpenAI API docs, Prompt engineering guide (message roles)
- Anthropic Claude docs, Release notes: System prompts
- Anthropic Claude docs, Prompting best practices (giving Claude a role)
- OpenAI (2023), chatml.md, Chat Markup Language documentation in the openai-python repository (created 1 March 2023)