Language models & chatbots
A system prompt is hidden instruction text set by the app developer that tells an AI model how to behave, before the user's own message is added.
When you chat with an AI product, your message is not the only thing the model sees. The developer has placed a system prompt ahead of it: 'You are a helpful assistant for a bank. Only discuss banking. Be brief. Never give investment advice.' The model treats these instructions as its standing orders.
System prompts set personality, tone, rules, and background facts, such as today's date or the user's name. They are how the same underlying model can be a coding assistant in one app and a patient tutor in another. They are usually invisible to users, though people sometimes coax models into revealing them.
A system prompt is guidance, not a lock. Models generally follow it, but clever users can sometimes talk their way around it, which is called a jailbreak. Serious safety rules therefore live in training and guardrails as well, not only in the prompt.
A customer-support bot that politely refuses to discuss anything except the company's products is following a system prompt you never see.