02: System prompt

What this does

Separates a fixed behavioral rule (SystemMessage) from the user's actual question (HumanMessage), and sends both as a conversation list instead of a single string.

from langchain_ollama import ChatOllama
from langchain_core.messages import SystemMessage, HumanMessage

llm = ChatOllama(model="llama3.2:3b")

messages = [
    SystemMessage("You are a senior backend engineer. Always answer in 2 sentences or less, no fluff."),
    HumanMessage("What is LangChain?"),
]

response = llm.invoke(messages)
print(response.content)

Why a list, not a string

Part 1 passed a plain string to .invoke(). LangChain accepted it as shorthand and silently converted it into a single HumanMessage behind the scenes. Here, passing an explicit list is what makes a multi-role conversation possible at all — a plain string has no way to express "this part is a rule, this part is the user."

The list is also the mechanism conversation memory is built on later: a real chat app appends each new HumanMessage and the model's AIMessage reply to the same list and resends the whole thing every call, because these models are stateless between requests — nothing is remembered server-side unless you resend it.

The three message roles

Type	Who writes it	Sent to the model as
`SystemMessage`	The application code	A `"system"` role message
`HumanMessage`	The end user	A `"user"` role message
`AIMessage`	The model's own prior reply	An `"assistant"` role message, when replayed back in history

Under the hood, LangChain converts these into the role-tagged JSON format every chat API actually expects ({"role": "system", "content": "..."}, etc.). The SystemMessage/HumanMessage classes exist so your code reads clearly and stays provider-agnostic — you're not hand-building that JSON or memorizing each provider's exact field names.

Order and position matter

The system message is conventionally placed first in the list. Most providers treat position as meaningful — a system instruction appearing after user turns is weaker or ignored by some models, because it reads more like "the user said this" than "this is a standing rule." Put SystemMessage first unless you have a specific reason not to.

Production considerations (not visible in this toy example)

Cost: the system prompt is resent on every single call in a conversation — it is not "set once." A long system prompt multiplies token cost across every turn.
Not a security boundary: a system prompt is a strong suggestion, not an enforced rule. It can be overridden or leaked by adversarial user input ("prompt injection") — never put secrets in a system prompt, and never treat it as a guarantee for safety-critical behavior.
Model-dependent strength: smaller/local models (like llama3.2:3b here) follow system instructions less reliably than larger hosted models. If output doesn't match the instruction, that's a real, known limitation — not necessarily a bug in the code.

One-line interview answer

"A system prompt is a SystemMessage placed first in the message list — it sets a standing rule the model should follow for the whole conversation, separate from what the user typed. It's not a security boundary; it's a strong instruction that a determined adversarial prompt can still override."