System prompt

Instructions given to the model outside the conversation, which apply to every turn rather than one message.

A system prompt sets standing rules: who the model is acting as, what it must always or never do, and what shape the output takes. It is passed separately from user messages, and models are trained to weight it more heavily and to keep following it as the conversation moves on.

The practical test for whether something belongs in a system prompt: would you otherwise paste it into every message? Role, tone, output format, domain constraints and refusal rules all pass. The actual task does not.

Different products expose this differently - custom instructions in ChatGPT, a system parameter in the API, a CLAUDE.md or AGENTS.md file in a repository - but the mechanism is the same.

Where this comes up

Prompts, configs and tutorials in the library that touch this term.