System prompt

A system prompt is the persistent instruction set given to an LLM that defines its role, behavior, constraints, output format, tone, and any guardrails. It's sent with every request and acts as the "operating system" the model runs against.

Distinction from user messages:

• System prompt. "You are a senior customer support agent at Acme. You answer technical questions about our API. You only answer from official docs. If unsure, escalate. Tone: friendly but precise."
• User message. "How do I authenticate to your API?"

The system prompt is durable. The user message is the immediate request.

Anatomy of a production-grade system prompt:

• Role definition. Who the agent is.
• Task scope. What it should do.
• Constraints + boundaries. What it should NOT do.
• Output format. How responses should be structured.
• Examples. Few-shot demonstrations of good outputs.
• Tone + voice. How it should sound.
• Escalation rules. When to defer to a human.
• Tool documentation. What tools it has access to + when to use each.

Good system prompts in production AI automation are typically 500-3,000 words. They're the highest-leverage code in any AI system — small changes in prompt design produce dramatic changes in output quality. This is why prompt engineering is a real discipline, not a casual activity.