System Prompts — Controlling Model Behavior

6.0 What This Teaches

• What a system prompt is and where it sits in the API call
• How to set a persona that changes tone and detail level
• Enforcing output format constraints across multiple turns
• The difference between system-level and user-level instructions
• A full working example using a code reviewer persona

6.1 What a System Prompt Is

6.2 Setting a System Prompt

# system_basic.py - compare output with and without a system prompt.
import anthropic

client = anthropic.Anthropic()

user_message = "Write a function that reads a JSON file."

# Without system prompt - model chooses its own style.
r1 = client.messages.create(
    model="claude-sonnet-4-6",
    max_tokens=256,
    messages=[{"role": "user", "content": user_message}]
)
print("--- No system prompt ---")
print(r1.content[0].text)

# With system prompt - model follows the constraints.
r2 = client.messages.create(
    model="claude-sonnet-4-6",
    max_tokens=256,
    system=(
        "You are a senior Python developer. "
        "Always include type hints. "
        "Return only code with no prose."
    ),
    messages=[{"role": "user", "content": user_message}]
)
print("--- With system prompt ---")
print(r2.content[0].text)

6.3 Persona Setting

# persona.py - two personas, same question, different outputs.
import anthropic

client = anthropic.Anthropic()

question = "What is a pointer?"

expert_system = (
    "You are a C++ expert focused on performance and undefined behavior. "
    "Assume the reader is an experienced systems programmer."
)

beginner_system = (
    "You are a patient teacher explaining programming concepts to beginners "
    "who have no prior experience with low-level languages. "
    "Use analogies and avoid jargon."
)

for label, system in [("Expert", expert_system), ("Beginner", beginner_system)]:
    r = client.messages.create(
        model="claude-sonnet-4-6",
        max_tokens=256,
        system=system,
        messages=[{"role": "user", "content": question}]
    )
    print(f"--- {label} persona ---")
    print(r.content[0].text)
    print()

6.4 Output Format Constraints

# json_system.py - enforce JSON output via system prompt.
import anthropic
import json

client = anthropic.Anthropic()

system = (
    "Always respond with valid JSON. No other text before or after the JSON. "
    "Every response must have exactly two keys: "
    "'answer' (a string) and 'confidence' ('high', 'medium', or 'low')."
)

questions = [
    "What does the `with` statement do in Python?",
    "Is Python dynamically typed?",
]

for q in questions:
    r = client.messages.create(
        model="claude-sonnet-4-6",
        max_tokens=256,
        system=system,
        messages=[{"role": "user", "content": q}]
    )
    data = json.loads(r.content[0].text)
    print(f"Q: {q}")
    print(f"A: {data['answer']}")
    print(f"Confidence: {data['confidence']}")
    print()

6.5 Scope: System vs User

6.6 Example: Code Reviewer Persona

# code_reviewer.py - system prompt makes Claude act as a strict code reviewer.
import anthropic

client = anthropic.Anthropic()

system = """\
You are a strict Python code reviewer. When shown code, you always check for:
1. Security issues (e.g., shell injection, hardcoded secrets)
2. Missing type hints on function parameters and return values
3. Missing error handling for I/O operations
For each issue found, state the line number (if visible), the problem, and a fix.
If no issues are found in a category, say "None found."
"""

code_to_review = """\
def load_config(path):
    import json
    f = open(path)
    return json.load(f)
"""

r = client.messages.create(
    model="claude-sonnet-4-6",
    max_tokens=512,
    system=system,
    messages=[{"role": "user", "content": f"Review this code:\n\n{code_to_review}"}]
)

print(r.content[0].text)

6.7 Exercise

6.8 Common Mistakes

6.9 Key Terms