System Prompt

What is a system prompt?

Every message you type to an AI is a "user message". But above all of it sits a rule you never see, one that's always in force — the system prompt. It sets the model's global rules and role: who it is, what tone to use, what it may do, and what it must not.

How is it different from a normal prompt?

It sits higher up
A user prompt is "the request for this turn"; the system prompt is "the constitution across all turns". It usually takes priority, and the model follows it first.
Only the developer sets it
Users generally can't change it. The product team writes it to define the assistant's persona and boundaries.
It applies everywhere
No matter what the user asks, the system prompt works in the background — everywhere, like air.

What goes into one?

Role definition
"You are a careful legal advisor" flips the model into professional language and tone.
Behavior limits
"Don't make things up" and "refuse illegal requests" are the safety floor.
Output format
"Reply in Spanish" or "list your points" keeps style consistent.
Tools and process
Telling the model which tools exist and how to call them is an agent's operations manual.

Why does it matter?

With the same base model, a well-written system prompt is the difference between a reliable assistant and one that rambles. It's also the top target of prompt injection — because whoever can rewrite it can bypass every limit.

Bottom line: the system prompt is an AI's invisible constitution — it decides the model's identity, boundaries and bottom line.