What is a system prompt?

The chat window on a store's website only wants to talk about orders and returns. Ask it for a poem and it politely steers you back to your package. Meanwhile, a general chatbot in another app is happy to write poems all day. They might even run on the same kind of model. The difference is a set of instructions someone wrote before you typed anything, called a system prompt.

A system prompt is a set of standing instructions for an artificial intelligence (AI) model. The company that builds the app usually writes it, and it can cover the model's role, its tone, what to help with, and what to decline. Anthropic's guide for its Claude models says that setting a role there focuses how the model behaves and sounds, and that even one sentence makes a difference. A shoe shop's might begin: "You are a helpful assistant for an online shoe store."¹

You don't see it, but it's there on every turn. Each time you send a message, the app sends the model the system prompt along with the whole conversation so far. Anthropic's documentation lists the system prompt among the things that fill the model's context window, the working space it reads before each reply. So the model reads those instructions again before it answers you, every time.²³

That's why one model can feel like several different products. A homework helper might be told to guide a student through a problem step by step instead of handing over the answer. A support assistant is told to stick to the company's policies. A coding tool is told to reply in code. Even the voice can change: one app asks for formal answers, another for a playful tone. The engine underneath can be the same. The instructions on top change, and so does what you get.

Many chatbots let you write standing instructions of your own, often called custom instructions or personalization. You can say who you are, what you're working on, and how you like answers to look. OpenAI's help pages describe its version as applying across your conversations until you change or remove it, so you don't have to repeat yourself every time.⁴

A system prompt guides the model. It doesn't lock it. A determined user can sometimes talk a model out of its instructions, or coax it into repeating them word for word. Security guidance from the Open Worldwide Application Security Project says a system prompt shouldn't be treated as a secret, or used as a security control. Real limits, like who can see your order history, have to be enforced by the app itself.⁵

If your chatbot has a place for custom instructions, try one standing line: "Keep answers short, and ask me a question if my request is unclear." Then notice whether the replies change. If they don't suit you, edit the line.