
TLDR
A system prompt is a hidden set of instructions given to an AI model at the start of a session that shapes its behavior, tone, and capabilities throughout the entire conversation.
Unlike a regular message you type into ChatGPT or Claude, a system prompt is set by the developer or operator of the application before any user interaction begins. The AI reads and follows these instructions throughout the conversation, but in most products the system prompt is invisible to the end user.
System prompts are used to define the AI's persona, restrict what topics it can discuss, set the language and format of its responses, and give it background context it needs to be useful. A customer support chatbot, for example, is typically built by wrapping a general-purpose AI like Claude or GPT-4 in a system prompt that tells it to act as a support agent for a specific product, stay on-topic, and respond in a certain tone.
For prompt engineers and developers, writing effective system prompts is one of the highest-leverage skills in AI work. A well-written system prompt can turn a general AI into a highly specialized tool. A poorly written one leads to inconsistent behavior, off-topic responses, or a model that ignores key constraints.
Customer support agent
A system prompt that tells the AI to act as a support agent for a SaaS product, respond only to product-related questions, always offer to escalate to a human, and reply in a friendly but professional tone.
Writing assistant with brand voice
A system prompt that defines a company's tone of voice, lists words to avoid, and instructs the AI to always write in second person and end emails with a specific sign-off.
Coding assistant
A system prompt that tells the AI to always respond with working code, explain each step, default to TypeScript, and never suggest deprecated APIs.
A regular prompt is what you type in a conversation with an AI. A system prompt is a separate, higher-level instruction that runs before any conversation starts. System prompts define the rules the AI follows for the entire session. Regular prompts are individual requests within those rules.
Usually not. In most AI products and APIs, the system prompt is hidden from end users. However, a sufficiently persistent user can sometimes extract parts of a system prompt by asking the model directly, which is why critical security rules should not rely on system prompts alone.
Be specific about the AI's role, tone, and constraints. Define what it should do and what it should not do. Include examples of good responses where helpful. Test edge cases: how should it handle questions outside its scope? Iterating on system prompts is a core skill in AI product development.
Yes. All major AI models support system prompts via their APIs. When you use consumer products like ChatGPT or Claude.ai directly, the companies themselves set a system prompt behind the scenes. When developers build on top of these models, they write their own system prompts to customize the AI's behavior.
Bottom line
A system prompt is a hidden set of instructions given to an AI model at the start of a session that shapes its behavior, tone, and capabilities throughout the entire conversation.
Prompt packages that apply these concepts directly.