System Prompts

Introduction

The system prompt sets the agent's persona, rules, constraints and tool contract. It's the most important prompt you'll write — it runs on every call.

Beginner analogy: the employee handbook every staffer reads on day one and keeps in their pocket.

Understanding the topic

Core concepts:

• Always sent — counts against context for every turn.
• Place identity, rules, formatting and tool docs here.
• Avoid putting user data in the system prompt unless intentional.
• Use prompt caching to make system prompts effectively free.
• Keep it stable; A/B test changes carefully.

Syntax reference

Visual workflow / architecture:

┌────────────────────────────┐
│  System Prompt             │
│  • Role                     │
│  • Tools (JSON schema)      │
│  • Output format            │
│  • Safety rules             │
└──────┬─────────────────────┘
       │
       ▼
   User turn 1, 2, 3 …

Real-world use

ChatGPT custom GPTs, Claude projects, Cursor's hidden system prompt — all rely on long, careful system messages.

Best practices

• Version the system prompt and link it to commits.
• Use prompt caching to slash cost.

Common mistakes

• Stuffing the system prompt with everything — context bloat.

Hands-on exercise

Interview preparation — practice these questions:

• Q1. Why does the system prompt run on every call?
• Q2. How does prompt caching save money?
• Q3. Scenario: your agent ignores its safety rules. What do you check first?