Before you say a word, the AI has already been told who it is and how to behave. Those first instructions are the system prompt.
On your first day at a new job, before any customer walks in, you read your job description: what the role is, how to speak to people and what you must never do.
A system prompt is that job description for an AI. It’s a set of instructions the AI reads before your conversation, setting its role, tone and rules. You usually don’t see it.
What goes in one
- who the AI is: “You are a friendly support assistant for a bike shop”
- how to speak: short answers, plain English
- what to avoid: no medical advice, no promises about refunds
- useful background, like opening hours or today's date
Why it matters
The same LLM can be a cheerful tutor in one app and a strict legal assistant in another. The difference is mostly the system prompt.
What it isn’t
It isn’t a lock. A clever user can sometimes talk an AI out of its instructions. That’s called prompt injection. It’s why guardrails exist too.
Where the comparison breaks down
A new employee remembers their job description all day. An AI re-reads the system prompt with every single message.
How the words fit together
The system prompt comes first. Then your prompt. Guardrails add checks around both.
When a company builds a chatbot, most of its personality and rules come from the system prompt, not from a different AI.

When you'll hear it
“Put it in the system prompt.” Add it to the standing instructions the AI reads before every chat.
“It ignored the system prompt.” The AI didn't follow its standing instructions.
“The system prompt leaked.” Someone tricked the bot into revealing its hidden instructions.
“What instructions does our AI assistant run with? Who's allowed to change them?”


