Prompt Templates

The problem with Lessons 1 and 2

So far, every message we sent was written directly into the code:

model.invoke("In one sentence, what is LangChain for?")

That's fine once. But real programs need to:

• Give the AI standing instructions (a personality, a rule like "keep answers short") without retyping those instructions into every message.
• Send many different questions that share the same instructions, where only a small part of the message actually changes each time.

A prompt template solves both.

A template is a message with a blank

prompt = ChatPromptTemplate.from_messages(
    [
        ("system", "You are a concise research assistant. Answer in at most two sentences."),
        ("human", "{question}"),
    ]
)

If you've used Python f-strings (f"Hello {name}"), this is the same idea: {question} is a placeholder, not a real question yet, it gets filled in later.

Two kinds of entries here:

• "system": standing instructions for how the AI should behave. Fixed, never changes between requests.
• "human": the user's actual message. Here it's a placeholder, waiting to be filled in.

Building this template is a one-time cost. From then on, "be a concise research assistant" applies automatically to every question you run through it, without repeating the instruction.

Filling in the blank

filled_prompt = prompt.invoke({"question": "What is LangChain for?"})

{"question": "..."} is a dictionary: a label ("question") paired with a value. LangChain matches that label against the {question} placeholder in the template and pours the value in. The label and the placeholder name must match exactly. Typo it ("query" instead of "question") and LangChain raises an error, it has no way to know what fills the blank.

filled_prompt isn't text yet, it's a small object holding the finished list of messages. Calling .to_messages() on it shows you exactly what will be sent:

[system] You are a concise research assistant. Answer in at most two sentences.
[human] What is LangChain for?

Sending it to the model

response = model.invoke(filled_prompt)

Same .invoke() from every earlier lesson. The only difference: instead of a single string, we're handing it a filled-in template, a full list of messages (system + human).

Why two separate steps, instead of one?

You'll notice this lesson calls prompt.invoke() and then model.invoke() as two distinct lines, rather than connecting them with | like you might have seen in tutorials. That's deliberate: Lesson 6 introduces | as its own concept. Here, the goal is just to understand what a template does, on its own, before adding anything else on top of it.