AI Engineering Tutorial · LLM Basics · Lesson 3 of 25

Messages, Roles and System Prompts

How a chat request is structured: the system prompt, user and assistant messages, and how conversation history works.

Create a free account to save this page, take notes and track your progress. Sign in

The message list

A chat request contains:

A system prompt: instructions and context that apply to the whole conversation (role, rules, tone).
A list of messages, alternating user and assistant, ending with the user's latest message.

The model reads all of it and writes the next assistant message. To continue a conversation, append the reply and the next user message, and send the whole list again.

A two-turn conversationTry it yourself

class StubModel:
    # Stands in for an LLM API so the example runs in the browser without a key.
    def __init__(self, replies):
        self.replies = list(replies)
        self.calls = []
    def create(self, system, messages, max_tokens=200):
        self.calls.append({"system": system, "messages": messages, "max_tokens": max_tokens})
        return self.replies.pop(0)

model = StubModel(["Order 10042 shipped yesterday.", "Yes, until it leaves the depot. Use Orders > Change address."])
system = "You are the support assistant for Booknook. Be brief. Never promise refunds."
messages = [{"role": "user", "content": "Where is order 10042?"}]
reply = model.create(system, messages)
messages.append({"role": "assistant", "content": reply})
messages.append({"role": "user", "content": "Can I still change the address?"})
reply = model.create(system, messages)
for m in messages + [{"role": "assistant", "content": reply}]:
    print(f"{m['role']:9}: {m['content']}")
print("messages sent on the second call:", len(model.calls[1]["messages"]))

class StubModel:
    # Stands in for an LLM API so the example runs in the browser without a key.
    def __init__(self, replies):
        self.replies = list(replies)
        self.calls = []
    def create(self, system, messages, max_tokens=200):
        self.calls.append({"system": system, "messages": messages, "max_tokens": max_tokens})
        return self.replies.pop(0)

model = StubModel(["Order 10042 shipped yesterday.", "Yes, until it leaves the depot. Use Orders > Change address."])
system = "You are the support assistant for Booknook. Be brief. Never promise refunds."
messages = [{"role": "user", "content": "Where is order 10042?"}]
reply = model.create(system, messages)
messages.append({"role": "assistant", "content": reply})
messages.append({"role": "user", "content": "Can I still change the address?"})
reply = model.create(system, messages)
for m in messages + [{"role": "assistant", "content": reply}]:
    print(f"{m['role']:9}: {m['content']}")
print("messages sent on the second call:", len(model.calls[1]["messages"]))

Output

user     : Where is order 10042?
assistant: Order 10042 shipped yesterday.
user     : Can I still change the address?
assistant: Yes, until it leaves the depot. Use Orders > Change address.
messages sent on the second call: 3

Good system prompts

Say who the assistant is and who it serves.
State rules that always apply ("answer only about our products", "never promise refunds").
Describe the output format if it matters.

Keep facts that change per request (the user's order, retrieved documents) in the messages, not baked into the system prompt. See System prompts.

Test yourself

Where should a rule like 'never promise refunds' go?

In the system prompt
In the first assistant message
Nowhere; models know it
In the max_tokens setting

Show answer

Answer: In the system prompt

The system prompt holds instructions that apply to the whole conversation.

Was this page helpful?