Q1. What is a system prompt in LLM chat interfaces?

A system prompt is an initial instruction that sets the overall behavior, persona, and rules for the assistant in a chat model.
It is separate from user messages and is not visible to the end-user in typical applications.
The system prompt influences the model's responses across the entire conversation.
Example: "You are a helpful coding tutor. Never provide complete code solutions; instead, guide the user with hints."
Unlike user prompts that change per turn, the system prompt stays constant.
It is a powerful tool for enforcing safety, style, and role consistency.

Q2. How does a system prompt differ from a user prompt?

System prompt sets the global context and rules for the assistant, typically sent once at the beginning.
User prompt contains specific queries or instructions, changing each turn.
The model is trained to follow system instructions with high priority.
Many APIs (OpenAI, Anthropic) have dedicated system message fields.

Q3. What are the best practices for writing effective system prompts?

Best practices:
• be clear and specific about the assistant's role and limitations
• set rules for formatting and tone
• use directives like "Do not..." and "Always..."
• keep the system prompt concise
• test with edge cases
• combine with few-shot examples for complex behavior
• iterate.
Example: "You are a medical information assistant. Provide general health information but not specific diagnoses. Always recommend consulting a doctor for serious symptoms.".

Q4. How can system prompts improve safety and reduce harmful outputs?

System prompts act as a first line of defense.
You can explicitly forbid generating hate speech, violence, or illegal content, instruct the model to refuse sensitive requests politely, set a persona that avoids controversial topics, require citing sources or indicating uncertainty, and add disclaimers.
Example: "Never provide instructions for creating weapons. If asked, respond: 'I cannot help with that request.'"
While not foolproof, a strong system prompt significantly reduces harmful outputs.

Q5. Can system prompts be combined with other techniques like few-shot and CoT?

Yes, system prompts work well with few-shot and CoT.
The system prompt sets the overarching behavior (e.g., "You are a logical reasoning assistant. Always show step-by-step reasoning.") while the user prompt contains the specific task and examples.
The system prompt ensures the assistant always uses CoT, even without being reminded each time.