System Instructions
System Instructions are a fixed instruction that a provider gives an AI text program before every conversation. They establish role, tone, and boundaries before the user has even typed anything.
Programs like ChatGPT answer questions typed into a text field. But what you type there isn’t all the program gets to read. Ahead of it sits another piece of text, written by the provider, which you never see yourself. This invisible preamble is called System Instructions. It specifies, for example, what role the program should play, how politely it should respond, and which topics it must not discuss. Think of it like the standing orders given to a new employee: they apply to every customer conversation, but no customer ever gets to see them.
The invisible frame around every chat
An AI model has no personality and no task of its own. It continues text, and it does so in the way it learned from enormous amounts of example text. Whether the result becomes a matter-of-fact assistant, a tutor, or a customer service agent is decided only by the instruction placed above it. This exact job is done by the System Instructions.
For companies, they are therefore an important control tool. Two firms can license the same model and still offer very different products. The difference often lies not in the model but in this preamble. Changing it takes minutes; retraining a model costs millions.
They are also the first line of defense on sensitive topics. A banking chatbot shouldn’t give investment advice, a school assistant shouldn’t fully solve homework. Such rules are typically placed in the System Instructions. They aren’t foolproof, though: users can partially bypass them with clever inputs. Experts call this prompt injection — smuggling in instructions through the text field.
Why the model weighs the instruction more heavily
Technically, System Instructions are simply text. They are passed to the model together with the user’s input as one continuous block, with a marker indicating which part comes from whom. The model thus reads both in a single pass. It has no separate memory for rules.
That the instruction nevertheless usually takes precedence is due to training. Models are specifically trained to weight the part marked as system higher than the user text. This hierarchy is a learned habit, not a built-in lock. That’s why it works in most cases, but not in all.
A practical side effect: the instruction takes up space in the context window, the limited amount of text the model can take in at once. A very long preamble with hundreds of rules therefore costs space, and with every request, money too. Developers thus try to phrase it as concisely as possible and test many variants against each other.
From chatbot templates to leaked preambles
Anyone building something with AI themselves runs into the term immediately. In the developer tools from OpenAI, Google, or Anthropic there’s a dedicated field for it. Even in no-code builders for custom assistants, the preamble hides behind fields like “role” or “behavior.” Even free chat interfaces often allow you to save your own instructions permanently.
In the news, System Instructions usually come up when they become public. Users repeatedly manage to get the hidden text revealed to them. Such leaks show how a provider actually steers its AI, and they regularly spark debates about censorship or political bias. Anthropic now voluntarily publishes the instructions for Claude.
The term shouldn’t be confused with a regular prompt. The prompt is a single question from a user and applies only to that moment. The System Instructions apply to all users and all conversations at once. And they are no substitute for training: whatever a model fundamentally cannot do, no instruction will teach it.