The direct answer
A system prompt is a set of instructions given to an AI before the conversation starts — invisible to you, but shaping everything it says. It defines the AI's role, its tone, its rules, and often basic facts it should know, like the current date.
Anthropic's own documentation says it plainly: the Claude app uses a system prompt to provide up-to-date information at the start of every conversation and to encourage certain behaviors. When developers build with the API, they write their own system prompt instead. Either way, there is always one — the AI never starts with a blank slate.
How it works
Think of a conversation with an AI as having layers. The system prompt is the bottom layer: the highest-priority instructions. Your message sits on top of it. When the two conflict, the system prompt usually wins — that's how an app can make its AI always answer in a certain format even if you ask for something else.
Who writes it depends on the product. App makers write system prompts for their users, which is why ChatGPT, Claude, and Gemini each have a distinct "personality." Developers using an API pass their own system prompt with every request. Even power users get a version of this: the custom-instructions settings in chatbot apps are just a system prompt with your name on it.
One analogy to hold onto: the system prompt is like the briefing an actor gets before going on stage — the character, the rules of the scene, what to avoid. It doesn't change the actor's talent; it directs how that talent gets used.
A simple example
As an illustration, imagine a pizza shop's AI assistant. Its system prompt might say: "You are a friendly assistant for Tony's Pizza. Keep answers under three sentences. Never invent menu items — only mention items from the menu list. If asked about anything else, politely redirect to pizza."
Two customers ask the same question — "Can you help me?" — and both get answers shaped by those invisible rules: short, friendly, and stubbornly about pizza. Neither customer ever sees the instructions that produced their answer.
Why it matters
System prompts explain small mysteries. Why do two AI apps answer the same question differently? Different system prompts. Why does an AI refuse some requests, or always respond in a certain format? System prompt. When a company "updates the AI's personality," they usually just rewrote this file.
They also explain prompt injection — the trick where a user writes "ignore all previous instructions" to make an AI misbehave. That attack works precisely because the system prompt is the thing being attacked: the user is trying to overrule the hidden layer. Understand that one mechanism and you understand both how AI products are controlled and how they get broken.
The common misunderstanding
The big one: a system prompt is not training, and it's not knowledge. People sometimes treat it like a secret identity — as if the "real" AI is whatever the system prompt says. It isn't. The system prompt is just instructions layered on top of the model's trained abilities; it can't give the model facts it never learned or skills it doesn't have.
Write "you are a chess grandmaster" in a system prompt and you get an AI that talks like one — not one that plays like one. Instructions shape behavior; training sets the ceiling. Confusing the two is how people end up both over-trusting AI ("it said it's an expert") and under-trusting it ("it's just following a script, so it can't reason").
What changed recently
System prompts have gone from secret sauce to published material. Anthropic now publishes the system prompts used by its consumer apps in its release notes — anyone can read exactly how Claude is briefed, and watch those instructions evolve over time. That's a remarkable shift from an industry that once treated these files as trade secrets.
They've also become the main craft of AI tuning. As models improve, more of the work has shifted from training tricks to clever briefing — longer, more structured system prompts that set roles, rules, and behavior in detail. The system prompt is where that craft lives now.
Try it on PlainLogic
The AI Lab on PlainLogic shows how instructions steer model behavior — and the Prompt Playbook walks through prompt structure, including the hidden layer that sits underneath everything you type.
Sources
- PRIMARY SOURCEAnthropic Docs: System prompts (published Claude system prompts)
- PRIMARY SOURCEAnthropic Docs: Prompt engineering overview