Custom Instructions

Anyone who works regularly with a chat program like ChatGPT quickly finds themselves repeating things. You add every time that the answer should be short, in English, and without bullet points. Custom Instructions take this repetition off your hands. You enter your preferences once in the settings, and the program then takes them into account in every new conversation. You can also provide information about yourself there, such as your profession or grade level. This is meant to make responses fit better right from the start.

What it saves you

The obvious benefit is time. Anyone submitting twenty queries a day would otherwise type the same three lines of instructions twenty times. These lines also cost computing effort, since every word of a request has to be processed. A text stored once is more convenient, even though technically it is still sent along with every request.

More important, however, is the quality of the responses. A language model knows nothing about the person in front of the screen. It has to guess how much prior knowledge someone has. If it states there that you are studying medicine, explanations about medications turn out differently than for a layperson. The exact same question then receives a different answer.

For companies, this is one building block of standardization. If all employees have entered the same instructions, texts sound consistent to the outside world. It is no substitute for real oversight, but it is a cheap first step.

Why the instructions don’t always take hold

Technically, nothing magical happens. The stored text is invisibly placed at the beginning of every new conversation, before the actual question follows. The model therefore sees a longer text than the one the user actually wrote. You can picture it like a note someone attaches to every task before passing it along.

This leads to an important limitation. A language model doesn’t follow instructions the way a computer program follows a rule. It merely weights them more heavily than other parts of the text. If the Custom Instructions say “maximum three sentences,” you sometimes still get five. The longer a conversation goes on, the more the initial note fades compared to what has been discussed since.

Custom Instructions should not be confused with the so-called system prompt. That one comes from the provider, is invisible to users, and contains basic rules, such as which content is refused. The user’s own instructions rank below that in priority. So they cannot override the provider’s rules. Nor are they a form of training: the model itself does not change as a result of the input.

From settings menus to company assistants

You most directly encounter the term in ChatGPT’s settings. There are two fields there: one for details about yourself, one for the desired response style. Other providers call the same thing something different, such as “Preferences” or “System Instruction.” Programming interfaces also have such a field, through which developers can set fixed instructions.

A typical everyday example: a student notes that he is in tenth grade and prefers explanations with everyday examples. After that, answers about chemistry or history automatically turn out more fitting. A programmer instead enters which programming language he uses, and never has to mention it again.

In the business press, the idea appears under other names. When a company offers its own assistant for customer service, what’s often behind it is nothing more than detailed fixed instructions. This is considerably cheaper than retraining a model. Part of what gets marketed as “our own AI” is exactly that.