System Prompts¶
System prompts are the primary way to define an agent's personality, expertise, and constraints. They are the foundation of agent identity — when we build multi-agent systems later, each agent gets a different system prompt that defines its role.
How System Prompts Work¶
The system role message is always the first message in the messages list. It sets the model's behavior for the entire conversation:
messages = [
{"role": "system", "content": "You are a senior financial advisor. Be precise and professional."},
{"role": "user", "content": "Should I invest in index funds?"},
]
The model treats the system prompt as its identity — it shapes tone, expertise, format, and constraints for all subsequent turns.
Same Query, Different Personas¶
The same user query with different system prompts produces completely different responses:
# Formal advisor
{"role": "system", "content": "You are a senior financial advisor. Be precise and professional."}
# Casual buddy
{"role": "system", "content": "You are a chill friend who knows about money. Keep it casual."}
Both receive the same question, but the first responds with structured recommendations and caveats, while the second gives relaxed, conversational advice.
Best Practices for System Prompts¶
| Practice | Example |
|---|---|
| Define the role clearly | "You are a customer support agent for an e-commerce platform." |
| Set constraints | "Keep responses under 100 words. Never discuss competitor products." |
| Specify format | "Always respond with a numbered list of recommendations." |
| Include expertise | "You have deep knowledge of AWS infrastructure and Kubernetes." |
| Add guardrails | "If you don't know the answer, say 'I'm not sure' instead of guessing." |
System Prompts in Multi-Agent Systems¶
In this workshop's multi-agent exercises, system prompts are what differentiate agents:
- Sequential pattern — each pipeline stage has a specialized system prompt (researcher, writer, editor)
- Group chat — the same shared conversation, but each agent's system prompt is prepended before its turn
- Handoff — the triage agent and specialist agents have different system prompts defining their scope
Temperature for system prompts
Pair system prompts with appropriate temperature settings. A formal advisor works best with low temperature (0.0–0.3), while a creative brainstormer benefits from higher values (0.7–1.0).
Key Takeaways¶
- System prompts define agent identity — the same model becomes different agents with different prompts
- System prompts set tone, expertise, format, and constraints
- In multi-agent systems, each agent gets its own system prompt
- Pair prompts with appropriate temperature settings for best results
Hands-On Exercise¶
Now try it yourself — head to the System Prompts exercise to see how the same query gets different responses with different personas.
You can run exercises from the terminal or use the Workshop TUI.