Introduction to System Prompts: What They Are and Why They Matter

What is a system prompt?

A system prompt is an instruction you give to an AI model that shapes how it behaves for an entire conversation. It's separate from the user's message — while the user says "solve this problem," the system prompt says "you are a helpful code reviewer who prioritises security and performance."

Think of it as the model's role, personality, and guidelines all in one. Every response the AI gives is filtered through the system prompt's constraints and instructions.

Why system prompts matter

A well-written system prompt is the difference between an API you struggle with and one that just works.

System prompt vs user message

Prerequisites

• Basic familiarity with what an LLM is (ChatGPT, Claude, etc.)
• Experience using an AI chat interface
• No coding required — all examples are prompt text

Note: This series is about crafting prompts, not code. If you can write clear English, you can write great system prompts.

What you will learn

By the end of this series you will be able to:

• Understand the three core parts of any system prompt
• Write clear, specific instructions the AI will actually follow
• Create prompts for different tasks: code review, customer support, content writing, data analysis, and more
• Design safety guardrails and constraint prompts
• Manage system prompts in production
• Recognize common pitfalls and anti-patterns
• Master advanced techniques like meta-prompting and dynamic prompts

Series overview

Here is everything covered across the 10 pages:

How system prompts work under the hood

When you call an LLM API, your request is structured as a list of messages — typically a single system message followed by a back-and-forth of user and assistant turns. The model reads the entire sequence top-to-bottom each time, so the system prompt isn't "remembered" by the server; your client re-sends it on every request. This is why teams often treat the system prompt as a static asset versioned in source control rather than as conversational state.

Under the hood, instruction-tuned models like Claude, GPT-4, and Gemini are trained to weight the system message more heavily than later user turns. That's what makes "you are a careful code reviewer" stick across dozens of follow-up messages without you having to repeat it. The downside: a long system prompt also consumes tokens on every request, which adds up at scale. A 1,200-token system prompt across one million requests costs roughly 1.2 billion input tokens — material in any production budget.

Treat the system prompt as the model's bias, not its policy. It's strong enough to shape style, tone, and reasoning patterns, but it is not a security boundary. If a user can send arbitrary input, assume a determined attacker can talk around any instruction you put in the system message. For real guardrails — secrets handling, rate limits, tool authorization — enforce them in code or via dedicated prompt security and guardrails patterns.

System prompts in popular APIs

Every major LLM SDK exposes the system prompt in roughly the same shape, with small naming differences. Here is the same prompt expressed in three providers.

from anthropic import Anthropic

client = Anthropic()
response = client.messages.create(
    model="claude-opus-4-7",
    max_tokens=1024,
    system="You are a senior Python reviewer. Prioritise security and readability.",
    messages=[{"role": "user", "content": "Review this snippet: ..."}],
)

from openai import OpenAI

client = OpenAI()
response = client.chat.completions.create(
    model="gpt-4o",
    messages=[
        {"role": "system", "content": "You are a senior Python reviewer. Prioritise security and readability."},
        {"role": "user", "content": "Review this snippet: ..."},
    ],
)

from google import genai

client = genai.Client()
response = client.models.generate_content(
    model="gemini-2.5-pro",
    config={"system_instruction": "You are a senior Python reviewer. Prioritise security and readability."},
    contents="Review this snippet: ...",
)

Claude uses a top-level system parameter, OpenAI uses a message with role: "system", and Gemini uses system_instruction. The string itself is portable across all three with minimal tweaking.

System Prompts FAQ

Start learning

Ready to go deeper? The next lesson covers system prompt anatomy — the three building blocks of every effective prompt. You can also jump ahead to persona and role-based prompting or explore the full tutorials library.

If you want the user-side counterpart, our prompt engineering introduction covers per-turn prompting techniques like few-shot examples and chain-of-thought. To see how system prompts power agentic behavior, jump to AI agents from scratch.