system prompt

in English: system prompt

basicwidespread

The system prompt is the background instruction set that defines how the AI should behave, what role to play, and what rules to follow before the conversation even starts.

In plain terms

Imagine a new colleague receiving an orientation sheet on their first day: who they work for, how to address clients, what they can and cannot discuss. The system prompt is the same for a language model. It is a text instruction not written by you but by the service creator, present in the background at the start of every conversation.

This is why the same model responds quite differently in a bank's customer service chat, a learning app for children, and a programming tool. The model is the same; only the guidance sheet differs. The system prompt determines the tone, role, prohibited topics, response format, and often which tools the model may use.

It differs from your own question in that you generally cannot see it, and the model gives it more weight than what you type. It is not an insurmountable wall: with clever questioning, the model sometimes deviates from it. It does not change the trained knowledge, only steers the behavior.

How it works

The service creator writes the instructions in everyday language: who the model is, what its task is, what style it should speak in, and what it should not do.
When you enter a question, the system inserts this text before your message in the background, so the model receives both together.
The model treats the system prompt as a high-priority instruction: it aligns its response to it, and if your request conflicts with it, it generally follows the system prompt.
The system prompt is re-injected before the model for every response because the model does not remember it on its own; therefore, its length also occupies space in the context window.
In the case of AI agents, the system prompt is much longer: it describes the available tools, the rules of the session, and when the agent must stop and ask for clarification.

An everyday example

In a webshop chat window, you ask what AI thinks about elections, and it politely states it can only assist with your order. The model did not come up with this on its own: the webshop's system prompt instructs it to stay on topic.

Who is affected

everyoneWhatever AI chatbot you use, its answers are shaped by an instruction invisible to you, so it is worth knowing that such instructions exist.

entrepreneur, executiveIf your company deploys its own AI assistant, the system prompt determines how it speaks to customers and what the company commits to on its behalf.

developerMost of an application or agent's behavior can be set by writing and testing the system prompt, without retraining the model. → Code

parent, student, teacherThe safety guardrails of AI tools for schools and children are largely provided by the system prompt, which is useful but not infallible protection. → Education & Family

A common misconception

MythThe system prompt is a secret, unhackable set of rules that the model always follows.

In realityThe system prompt is only a strong guideline: with clever questions, the text can often be coaxed out, and the model can be induced to deviate from it. Therefore, passwords or secret data should not be written into it, and serious protection should not rely on this.

Limits and risks

You usually cannot see what was written into it, so you cannot be sure by what criteria the received answers are guided.
No guarantee: the model may ignore your instructions during long conversations or in response to tricky requests (this is one goal of prompt injection and jailbreaking).
It does not expand the model's knowledge nor eliminate its errors; the "do not make up data" instruction reduces but does not rule out hallucination.
An overly long or contradictory system prompt degrades response quality and adds extra cost to every message, which caching only mitigates.

Similar, but not the same

prompt (user prompt) – You write the prompt for a specific question, while the service provider sets the system prompt in advance, which applies to the entire conversation.
fine-tuning – Fine-tuning changes the model itself through further training, while the system prompt is only a text instruction that leaves the model untouched.
custom instructions – You set individual instructions in your own account; these are added alongside the provider's system prompt, not replacing it.

Related concepts

Broader

prompt, Prompting

Narrower

role assignment

On the edge

custom instructions, guardrail, prompt injection, context window, jailbreak, fine-tuning, AI-agent, large language model

A brief history

The term became known with chat language models: in the ChatGPT API opened by OpenAI in March 2023, messages were given a separate "system" role, and this division quickly became an industry standard. Since August 2024, Anthropic has publicly published the system prompts of its own chat applications, and since then, other developers have taken similar steps.

Where things stand

At the end of 2026, virtually every AI chatbot, corporate assistant and agent runs on a system prompt. For agents, this is often a carefully tested text of several thousand words, and writing it has become a professional task in its own right.

Status: 2026 November

Also known as

rendszerprompt · system prompt · system message · system message · system instruction · system instruction · developer message · system prompt

Sources

This entry was machine-edited and has undergone machine expert review; the “In the news” section updates automatically from Agent Age stories.