System prompt
The hidden standing instructions a company gives a chatbot before you say a word: who it is, what it’s for, what it won’t do. You usually can’t see it, but it’s in every conversation.
A system prompt is a block of instructions placed at the very start of a conversation, before the user’s first message. Whoever deploys the model writes it: the company behind a chatbot, the retailer running a shopping assistant, a developer building an app. It sets the persona (“you are a friendly beauty advisor”), the job, the house rules, and the topics to avoid.
To the model, it’s just more text in the context window , but fine-tuning teaches models to treat it as higher-ranking than what the user types. That’s what lets a company build a product on a general-purpose model: the same underlying LLM becomes a coding helper, a tutor, or a sales clerk depending on the sheet it was handed.
It’s also why two chatbots on the same model can behave so differently, and why it’s worth remembering that there’s usually a third party in your conversation whose instructions you can’t read. Most chat apps also let you add standing instructions of your own, which go in the same place.