The analogy

Imagine a shop assistant on the first day of work. Before the first customer comes in, the owner hands them a sheet: "always greet people, be formal, don't talk about politics, no discounts above ten percent, if you don't know something ask the storeroom". The customer, when they come in, never sees that sheet. But they feel its effects in every answer: the way of greeting, the topics the assistant avoids, the limit they won't cross.

The system prompt is that sheet behind the scenes. It's the first thing the AI "reads", before you, and it sets the character it plays and the rules it sticks to. When you notice the AI has a certain tone or refuses to do something, often it's that sheet talking.

How it really works

The system prompt is the top-priority block of instructions that the AI receives at the start of every conversation. It defines the assistant's identity, the way of answering, the forbidden topics and the safety rules. It's invisible to the user and carries more weight than your requests: when what you ask goes against those rules, as a rule the rules win.

Be careful not to confuse it with your personal instructions. The system prompt is written by the service provider and is the same for all users. Custom instructions, on the other hand, are written by you and apply only to your account: they're a second sheet, yours, that adds to the company's without being able to override it.

What you can do in practice

  • You can't modify the official system prompt, but you can add your own layer of rules with custom instructions, or by stating at the start of the chat the role and format you want ("act as a strict reviewer, answer in bullet points").
  • When the AI refuses or dances around a request, read it as the rule sheet at work, not as a whim: there's often a safety limit behind it.
  • If you build your own custom assistant, you write the system prompt: that's where you decide the character, tone and boundaries of the tool.

A common misconception

People believe the system prompt is "programming code", something technical and locked, written in a language for specialists. In reality it's written in the same language you speak: ordinary sentences, in Italian or English, like "you are a courteous assistant, don't give medical advice". The difference from your requests isn't the language, it's the position and the weight: it comes first and counts more. It's instruction, not code.

Frequently asked questions

Can I read ChatGPT's or Gemini's system prompt?

As a rule no: it's hidden on purpose. Sometimes an extracted or reconstructed version circulates online, but it isn't official material and it changes over time. To use AI well you don't need to see it: it's enough to know it's there and that it explains a lot of its behavior.

Is it the same thing as my prompt?

No. Your prompt is the request you write in the moment; the system prompt is the rule sheet the AI received earlier, valid for the entire conversation and with higher priority. Yours rests on top of its.

Can I write one of my own?

Yes, in two ways. With custom instructions you set personal rules valid for every one of your chats. And if you create a custom assistant (a personalized GPT or a tool via a platform), you write its system prompt from scratch, deciding its identity and limits.

Does the system prompt serve to "program the AI to deceive me"?

No, and it's a suspicious reading that leads you astray. Those instructions serve mostly to give consistent behavior and to keep safety limits, not to manipulate you. They're also anything but infallible: the AI sometimes applies them badly or works around them on its own. Seeing them as a conspiracy makes you lose sight of the useful thing, namely that many answers and many refusals are explained precisely by that initial sheet.