AI Prompt Engineering · Lesson

Injecting Persistent Behaviors

Rules that apply across all turns: always respond in JSON, never discuss X.

What Are Persistent Behaviors?

Persistent behaviors are rules that apply to every response the model generates, regardless of what the user asks. They are defined in the system prompt and never change during a conversation session.

Common persistent behaviors:

Always respond in JSON
Never discuss competitors
Always ask for clarification before writing code
Always cite sources
Always use a specific language or tone

Always Respond in JSON

Forcing the model to always return JSON makes output programmatically predictable. The system prompt must be explicit about this requirement:

import anthropic, json

client = anthropic.Anthropic(api_key='YOUR_API_KEY')

SYSTEM_JSON = '''
You must ALWAYS respond with a valid JSON object. No prose, no markdown, no code fences.
Every response must have at minimum: {"response": "string", "confidence": "high|medium|low"}
If you cannot answer, return: {"response": null, "confidence": "low", "reason": "string"}
'''

def ask(question):
    r = client.messages.create(
        model='claude-opus-4-5', max_tokens=300,
        system=SYSTEM_JSON,
        messages=[{'role': 'user', 'content': question}]
    )
    return json.loads(r.content[0].text)

result = ask('What is the capital of France?')
print(result['response'])    # Paris
print(result['confidence'])  # high

All lessons in this course

← Back to AI Prompt Engineering