Prompt Engineering — System vs User Prompts

• Use the system parameter to set role, tone, rules.
• Write clear, specific user prompts (and why specificity wins).
• Constrain output format (JSON, bullet list, word count).
• Control creativity with temperature.

Question: "What is gravity?"

system = "You are a pirate."        → "Arr, gravity be the force that...""
system = "You are a physics prof."  → formal, precise definition
system = "Explain to a 6-year-old." → "It's what makes things fall down!"

The system parameter

import anthropic
client = anthropic.Anthropic()

msg = client.messages.create(
    model="claude-haiku-4-5",
    max_tokens=300,
    system="You are a friendly Malaysian science tutor. "
           "Always use simple words and a local example.",
    messages=[{"role": "user", "content": "What is evaporation?"}],
)
print(msg.content[0].text)

The system prompt is a separate top-level parameter, not a message. It applies to the whole conversation.

Be specific — vague in, vague out

✗ "Write about dogs."
✓ "Write a 3-sentence fun fact about why dogs wag their tails,
   for a 9-year-old, ending with a question."

Constrain the format

system="Reply ONLY with valid JSON, no prose."
user="List 3 planets with their order from the sun as "
     '{"planet": ..., "order": ...} objects in an array.'

Telling the model the exact output shape makes the result programmable — you can json.loads the reply. (For guaranteed structure, tool use in Lesson 42 is even more reliable.)

temperature — creativity dial

# 0 = focused/deterministic; 1 = creative/varied
client.messages.create(model="claude-haiku-4-5", max_tokens=200,
    temperature=0,     # facts, code, classification
    messages=[...])

client.messages.create(model="claude-haiku-4-5", max_tokens=200,
    temperature=1.0,   # brainstorming, stories, names
    messages=[...])

Give an example or two

Showing the model the style you want (a sample input→output) steers it strongly. That's "few-shot prompting" — the whole of the next lesson.

# tutor.py — system prompt locks the behaviour
import anthropic
client = anthropic.Anthropic()

SYSTEM = (
    "You are 'CikguBot', a patient tutor for Malaysian students aged 10-14. "
    "Rules: use simple English; give one local everyday example; "
    "end with a short question to check understanding; "
    "never give the full answer to homework — guide instead."
)

def ask_tutor(question):
    msg = client.messages.create(
        model="claude-haiku-4-5",
        max_tokens=300,
        temperature=0.3,
        system=SYSTEM,
        messages=[{"role": "user", "content": question}],
    )
    return msg.content[0].text

print(ask_tutor("Why does ice float on water?"))

Sample output

Ice floats because it is lighter than the water around it — when water freezes
it spreads out a little and takes up more space, like how a packed bus feels
fuller than an empty one. That's why ais kacang ice cubes bob at the top of
your bowl! Can you think of another thing that floats because it's light for
its size?

Read the diff

The system prompt enforces the persona and rules for every question — simple words, a local example (ais kacang!), a check-question, and the "guide, don't solve" rule. The user only sends the topic. temperature=0.3 keeps it warm but consistent. This is how real LLM products behave reliably.

System prompt = role + rules (whole conversation); user prompt = the request. Be specific; constrain the format to make output programmable; use temperature 0 for facts/code and higher for creativity. Most "prompt engineering" is just clear instructions. Next: teach by example with few-shot prompting.