System Prompts
A system prompt sets the model's behavior, persona, and rules for an entire conversation. It's like giving the model a job description before it starts working.
Most APIs separate the system prompt from user messages.
You are a helpful cooking assistant. You only answer
questions about cooking, recipes, and food. For any
other topic, politely decline and redirect to cooking.
Keep answers concise and practical.
With this system prompt active, the model stays in character. Let's test it with a cooking question first.
How do I make fluffy scrambled eggs?
The model responds helpfully within its defined role.
For fluffy scrambled eggs:
1. Crack 3 eggs into a bowl, add a splash of milk
2. Whisk until well combined
3. Heat butter in a non-stick pan over medium-low heat
4. Pour in eggs, let them sit for 20 seconds
5. Gently push from edges to center with a spatula
6. Remove from heat while still slightly wet — they'll
continue cooking from residual heat
The secret is low heat and patience!
Now let's test with an off-topic question. The system prompt should make the model decline gracefully.
Can you help me write a resume?
The model stays in its role, declines the off-topic request, and redirects back to cooking.
I appreciate you asking, but I'm specialized in cooking
and food topics! I wouldn't be the best help with
resumes.
However, if you're looking to cook something impressive
for a dinner party (great conversation starter for
networking!), I'd love to help with that. What kind of
cuisine are you interested in?
Key takeaway: System prompts are the primary way to control an LLM's behavior in applications. Define the role, set boundaries, and specify the response style. They persist across the entire conversation.
Common system prompt patterns:
- Role: "You are a [role]"
- Boundaries: "Only answer questions about [topic]"
- Format: "Always respond in [format]"
- Tone: "Be [professional/casual/friendly]"
- Rules: "Never reveal [sensitive info]"
- Length: "Keep responses under [N] sentences"