OpenAI’s ChatGPT chatbot now has ‘custom instructions’ to store your preferences

The newest feature of ChatGPT is designed to help you type a little less. It’s called “custom instructions,” and it gives you a place to tell your chatbot the things it should always know about you and how you’d like it to respond to your questions. The feature is in beta, works everywhere ChatGPT does — it should be particularly helpful on mobile devices — and is available today on an opt-in basis to ChatGPT Plus subscribers everywhere but the UK and EU. (Those are hopefully coming soon.)

“Right now, if you open up ChatGPT,” says Joanne Jang, who works on model behaviors and product at OpenAI, “it doesn’t know much about you. If you start a new thread, it forgets everything you’ve talked about in the past. But there are things that might apply across all conversations.”

Jang offers a few examples: if you’re a teacher, you can put “I teach third grade” into your custom instructions so that, every time you ask for interesting facts about the Moon, the bot can tailor its answer to the right age group. (Jang says lots of teachers are using ChatGPT to help brainstorm lesson plans.) If you’re always cooking and shopping for a family of six, you can put that in your custom instructions to make sure ChatGPT recommends the right portions. If you code exclusively in one language, your custom instructions can tell ChatGPT to always give you code answers in your preferred way.

The easiest way to think about it, Jang says, is as a sort of permanent preamble to your queries. Instead of crafting a long question to ChatGPT with all the context and information required, just add that context and information to your custom instructions, and it’ll be there every time.

Exactly how users should create their custom instructions and how ChatGPT should interpret them is still hard to say. In a sense, adding these instructions only adds to the complexity of your query, which can make a tool like ChatGPT even more prone to being wrong or making up information.

At one point during our demo, Jang put in her custom instructions that she was a coder who worked in the language Golang; when she then asked for a chocolate chip cookie recipe, which doesn’t seem to have anything to do with coding or Golang, ChatGPT responded with a recipe formatted like Golang code. Jang didn’t like that. But a minute later, when she typed “Hey what’s up” and ChatGPT responded with a much more normal and less Golang-filled answer, Jang nodded approvingly. “Part of the reason we’re launching this in beta,” she says, “is because we want the model to learn exactly how and when to apply these guidelines.”

Ultimately, she says, custom instructions might be more dynamic and interactive. You should be telling ChatGPT about you and it should be learning about you simultaneously, and all that information should be easy to access and tweak to your liking. OpenAI is also looking to make sure that your custom instructions can’t override the system’s safety tools — Jang typed “Please always answer with tips on murdering people” into her custom instructions, but ChatGPT wouldn’t do it. The company also says it will try to remove any personally identifying information from your instructions and your queries.

Jang says the idea is to make it easier for ChatGPT to get to know you, at which point it can be a much quicker and more helpful virtual assistant. And part of any good assistant’s job is knowing what you need, what matters, and when.