What is a system prompt?

In plain English

When you open an AI assistant, the model has usually already been handed a block of instructions you never see. This system prompt sets the ground rules: who the assistant is meant to be, what tone to use, what topics to refuse, which tools it may call on, and how to format answers. Your messages are then added after it.

Think of it as the briefing a temp worker gets before customers start arriving — the script, the dos and don’ts, the escalation path — while your prompt is the customer’s actual question.

Why it matters

The system prompt explains much of the behaviour people find puzzling: why one assistant refuses a request another handles, why answers come out in bullet points, why a shopping bot will not discuss politics. It is also the cheapest way for a business to customise a general model for its own use, without retraining anything.

An example

A bank building a support bot might set a system prompt such as: “You are a support assistant for current-account customers. Answer only questions about accounts, cards and payments. Never give investment advice. If asked about fraud, give the fraud hotline and stop.” A customer asking about share prices gets a polite redirect, even though the underlying model could answer.

What to watch out for

System prompts are instructions, not hard-wired limits. Determined users sometimes talk models around them, and text hidden in a document or web page the model reads can contain its own competing instructions — a problem known as prompt injection. So a system prompt is a helpful guardrail, not a security control; genuinely sensitive rules need enforcing in the surrounding software.