User Prompt vs System Prompt

Introduction

Modern chat APIs split a prompt into system and user messages. The system message sets the model's role, rules and constraints — it's invisible to the end user. The user message is what the actual person types. Mastering this split is the first step from hobbyist to engineer.

Beginner analogy: The system prompt is the job description handed to a new employee on day one. The user prompt is the daily ticket from a customer. Same employee, very different behaviour based on the job description.

Understanding the topic

Core concepts to understand:

• System: persistent role, tone, rules, format — set once per conversation.
• User: per-message question or instruction from the human.
• Assistant: prior model replies (used to maintain conversation context).
• System prompts have higher 'weight' on Claude and GPT-4 — use them for guardrails.
• Sensitive instructions belong in system, never in user input.

Syntax reference

Visual workflow / architecture:

messages = [
  { role: "system",    content: "You are a polite Spanish tutor." },
  { role: "user",      content: "How do I say good morning?" },
  { role: "assistant", content: "Buenos días!" },
  { role: "user",      content: "And good night?" }
]

Real-world use

OpenAI's GPTs, Claude Projects, Cursor rules and Notion AI personas are all just persistent system prompts wrapped in a UI.

Best practices

• Put role, format and safety rules in the system message; leave the user free to ask anything.
• Keep system prompts under ~500 tokens for cost & clarity.
• Never trust user input to override the system — design defenses early (see Prompt Injection).

Hands-on exercise

Interview preparation — practice these questions:

• Q1. Difference between system and user prompts?
• Q2. Where do you put safety guardrails and why?
• Q3. What happens if a user message tries to override the system?
• Q4. Why split into roles at all instead of one big prompt?