Tutorials · Chapter D (4/4) · ~10 min

Talk to an LLM from code

Play → read → next

Shape a real LLM API call — system + user messages you can run and show.

Code Lab

Talk to an LLM from code

Practice message lists (system + user) like real APIs — offline fake client.

Python

# Fake LLM client — practice the shape of an API call
def chat(messages):
    user = messages[-1]["content"]
    return {"role": "assistant", "content": f"Echo coach: {user.upper()}"}

messages = [
  {"role": "system", "content": "You are a concise coach."},
  {"role": "user", "content": "Give me one study tip"},
]

reply = chat(messages)
print(reply["content"])

# Challenge: change the user message to ask for a focus tip, then run

Recap

What you just did

You composed an API-shaped prompt: roles (system/user), content strings, maybe temperature. You saw that the product is not “open a website” — it’s a request your program can repeat, log, and wrap in features.

Teach

How it works

Python-ish sketch of the shape:

messages = [
    {"role": "system", "content": "Reply in one short sentence."},
    {"role": "user", "content": "Name one use of embeddings."},
]
# client.chat(model="...", messages=messages)  → text

System steers job + style
User is the ask
Response text is what your app displays or pipelines next

Mental model: the LLM is a callable function — strings in, string out — not a person in a browser.

Use it

When you'd use this

Drafting helper bots inside your own app
Generating labels or summaries for a batch of rows
Prototyping before you add retrieval or tools

Watch out

Never bake API keys into shared notebooks or git. Responses can be wrong with full confidence — keep a human check for money, health, or legal. Cap length so a runaway reply doesn’t flood your UI.

Try next

Try this next

Change only the system message (same user ask). Save both replies. That’s prompt control you can show in two screenshots.