What are the three message roles in a chat style LLM prompt and what does each represent?

Chat style LLM (large language model) APIs (OpenAI Chat Completions, Anthropic Messages, Gemini, most open-source servers) take a list of messages. Each message has a **role**: `system`, `user`, or `assistant`. Many APIs add a `tool` role for tool results, and OpenAI's newer models use a `developer` role in place of `system`. The system role carries instructions that should hold across the whole call. They say who the model is, how it should respond, what format to use, and what it must refuse. You typically write this once per session and reuse it. The user role carries the actual question or task for this turn. In a chatbot it is whatever the human typed. In an internal pipeline it is whatever your code is asking the model to do. The assistant role holds the model's earlier replies in this conversation. Your code passes them back in on each new call so the model can see what it has already said. This is how multi-turn memory works on a stateless API such as Chat Completions (it remembers nothing between calls): you build the running transcript on your side and resend it each time.

There are three roles. The **system** role is the model's standing orders for the call: persona, output format, refusal behavior, tone, schema. It is set once and stays stable across turns. The user role is the current turn's input. In a customer-facing product it is whatever the user typed; in a backend pipeline it is whatever your application is asking. Retrieved RAG context and dynamic facts usually live here too, close to the question that needs them. The assistant role replays the model's previous responses in this conversation. APIs such as Chat Completions and Anthropic Messages are stateless: nothing is remembered between calls. To get coherent multi-turn behavior, your application stores the transcript. It resends user + assistant messages on every new request so the model has full context. OpenAI's Responses API can keep that state on the server instead, but the model still reads the whole history each call and you pay for it as input. The practical discipline is to keep stable rules in system and per turn content in user. Mixing the two is a common message-role mistake. Two more roles are worth knowing: a tool role carries tool results, and OpenAI's o1 and newer models take a developer message in place of system.