What are ChatGPT's new custom instructions, and how do they work?

Yesterday, July 20, 2023, OpenAI added a small setting to ChatGPT that changes how many people will use it. It is called custom instructions. Instead of opening every chat by explaining that you are a teacher, that you prefer short answers or that you cook for six, you write it once in your settings, and ChatGPT keeps it in mind from then on.

OpenAI says users told it that starting each conversation from scratch was a nuisance. The company also says it spoke with users in 22 countries about steerability, the ability to guide how a model responds, and found that people's needs and contexts vary widely. Custom instructions are its answer.

What OpenAI released

The feature is a beta, a test version that may change. It starts with ChatGPT Plus, the paid plan, and OpenAI says it will reach all users in the coming weeks. For now it is not available in the United Kingdom or the European Union.

In the settings, you find two boxes, each answering one question:

“What would you like ChatGPT to know about you to provide better responses?” This is for your background: your job, your audience, your location, your level of expertise.
“How would you like ChatGPT to respond?” This is for style and format: length, tone, structure, what to leave out.

TechCrunch reports that each box holds up to 1,500 characters, about 250 words, and that changes apply from your next conversation. It also notes that OpenAI had been quietly testing the feature with some users in the days before launch.

How it works

A chat model like ChatGPT has no memory between conversations. Each new chat begins empty, which is why people end up repeating themselves. Custom instructions fill that gap in a simple way: in effect, the text you write is added to the context the model sees each time it answers, much like a note pinned above every conversation. OpenAI says the model considers the instructions every time it responds.

OpenAI's own examples show the idea:

Who	What they write once	What changes
Teacher	Works on science programs for third-grade students; wants work options compared in a table of pros and cons	A question about the Moon gets three child-friendly topics, with a table weighing each
Developer	Writes only in Go; wants code without explanation, favoring efficient solutions	A request for a Fibonacci function returns short Go code instead of Python with commentary
Parent	Shops for a family of six	A pizza-night grocery list comes with quantities for six people

The instructions also reach plugins, the add-on tools that Plus users can enable. If you mention the city where you live and use a restaurant-booking plugin, ChatGPT may include that city when it calls the plugin.

Worked example: before and after

Here is a set of custom instructions a small-business owner might write.

Prompt

What would you like ChatGPT to know about you?
I run a two-person bakery in Portland, Oregon. I'm not technical. Most of what I write is for customers: social posts, emails and signs.

Prompt

How would you like ChatGPT to respond?
Use plain, friendly language. Keep answers under 150 words unless I ask for more. When you suggest options, give three at most. Use US spelling.

Now the owner types only "Write a post about our new rhubarb pie." Without instructions, ChatGPT might produce a long, generic text. With them, it should return a short, warm post written for local customers, with no need to restate the bakery, the audience or the length.

Why it matters, and what to watch

Custom instructions make a good prompting habit automatic. Giving ChatGPT context and a clear format has always improved answers; now that context is there by default. For regular users, it saves typing and makes results more consistent.

OpenAI is frank about the limits. During the beta, ChatGPT may sometimes ignore an instruction or apply it where it does not fit. Instructions that break OpenAI's usage policies may be refused when you try to save them, a check performed by its moderation system, and the model can also ignore them.

Privacy deserves a moment of thought. OpenAI says it may use custom instructions to improve its models, after removing personal identifiers, unless you turn this off in your data controls. Keep the boxes to what helps, and leave out anything you would not want stored.

Try it: Ask ChatGPT one question you ask often, such as "Plan a simple dinner for tonight." Save the answer. Then write two short custom instructions about your household and the style you like, ask the same question in a new chat, and compare the two answers line by line.

How do you write good custom instructions?

First, turn the feature on. On the web, Plus users click their name, open Settings, choose Beta features and opt into Custom instructions; the option then appears in the menu under your name. On the iPhone app, go to Settings, then New Features, and switch it on.

Then keep the first box factual and relevant: your role, your audience, your level of knowledge and any constraint that applies to most of your requests, such as a language, a country or a family size. Use the second box for rules about form: length, tone, structure and what to skip. Short, specific sentences work better than a long biography.

Finally, test and adjust. Ask a few typical questions and see whether the answers follow your instructions. If a rule is ignored, make it clearer; if it keeps showing up where it does not belong, narrow it, for example "for work emails" rather than "always." And remember that you can still give extra instructions in any chat; the saved ones set the default, and what you type in the moment adds to it.