MMOOZZGG.com

Glossary

What is a system prompt

In short

A system prompt is a standing instruction for an AI model that the developer sets before the conversation starts: who the AI is, what it may and may not do, and what tone to use.

Prompts · system prompt

MMOOZZGG.com editorsChecked: 1 min read

For a child

Before a school play, the teacher tells each child who they play and which rules to follow: stay in your role, speak politely, do not touch other people’s props. The audience does not hear these words, but the whole play follows them. An AI helper also has rules like these behind the scenes. They were written in advance, and they work for as long as the conversation lasts.

In plain words

When you write to a shop’s chatbot, it has already received a long note from its makers that you cannot see. The note says what the bot is called, what it talks about, what it keeps quiet about, what tone it uses and when it calls a person. That note is the system prompt. The model reads it before every answer, so the bot usually stays on topic even in a long chat. Your messages come after this note and do not replace it.

For business

For a company, it is the main way to set up an AI assistant’s behavior without programming or fine-tuning. It holds the brand’s tone, the limits of responsibility, the rules for personal data and the moment to hand the chat to an employee. A good instruction is short and testable: you can test each rule on real chats. A long text is sent with every request and raises the token bill. The main mistake is to treat this instruction as protection. Its content can leak, so keys and passwords never go there, and the assistant’s rights are limited in the systems themselves. Test the set of rules in a pilot before it goes to clients.

For a specialist

A system prompt is a message with the role “system” that an application places before the dialog when it calls a chat model’s API. It sets the role, the task, the limits, the answer format, tool descriptions, and service facts such as the date and the user. Unlike a user prompt, it is set once per session. Wallace et al. (2024) showed that models often give the system message the same priority as text from untrusted users and third parties. They proposed an instruction hierarchy: the model is trained to ignore less privileged instructions when they conflict. In its 2025 list, OWASP named system prompt leakage as a separate risk (LLM07). A system prompt is not a secret, it must not hold credentials, and authorization and separation of rights must run in deterministic code outside the model.

Example

An online shop gives its assistant a rule: answer only about orders and delivery, take prices from the catalog, and pass the chat to an operator if there is a complaint. A customer asks about the weather, and the assistant politely brings the talk back to the order.

Related terms

Questions and answers

Where is a system prompt written? In the settings of an AI assistant, or in the system field when you call a model’s API. The user usually does not see it, and the model gets it before every answer.
How is a system prompt different from an ordinary prompt? A user writes an ordinary prompt for one task. A developer sets the system prompt once, and it applies to the whole conversation: the role, the rules and the limits.

Sources

Where companies use it

Not sure whether your company needs this?

AI solutions map
Contact us