OpenAI's custom instructions feature that rolled out to ChatGPT Plus subscribers in July, is now available to all users.

The addition of custom instructions puts a new setting in your ChatGPT profile on desktop or iOS, and applies that setting across all of your conversations. Instead of having to give ChatGPT certain instructions at the beginning of every new conversation, it automatically calibrates its responses based on descriptions you add in settings. Just fill out the fields in ChatGPT settings, and see how it adapts.

The custom instructions section of settings contains two text entry fields. One asks what you'd like ChatGPT to know about you, and helpfully, (or creepily?) provides suggestions like "Where are you based?" and "What do you do for work?" The other asks how you'd like ChatGPT to respond, and asks questions like "How formal or casual should ChatGPT be?" and "Should ChatGPT have opinions on topics or remain neutral?"

It's an open-ended feature with infinite potential effects. As such, some applications for custom instructions are bizarre, and some are practical. Here's how people are making use of this new option:

One user posted a ChatGPT text output on X that they said seemed to "'tour' mildly disallowed behaviors [like] unprompted self-harm, nudity, [and] bio details of a (confabulated) non-celebrity," when they changed the setting to 1,500 repetitions of the letter "a" and an incomplete sentence (a known token repetition hack).

Tweet may have been deleted

Others are using custom instructions that are proving to make ChatGPT more accurate and effective.

Tweet may have been deleted

And then there's this custom instruction, which is quite the literary challenge.

• Researchers jailbreak AI chatbots, including ChatGPT
• ChatGPT has an Android app now
• OpenAI updates GPT-4 with new features
• OpenAI launches webcrawler GPTBot, and instructions on how to block it
• OpenAI announces $5 million partnership to support local news

Tweet may have been deleted

OpenAI also says it won't use instructions that violate its usage policies. And if you're wondering what OpenAI does with settings that might inadvertently reveal personal information about you, OpenAI says, yep, it might use that data unless you opt out of saving your chat history.

A collection of more generic presets with yes-or-no functionality, like "orient answers toward kids," or "assume user has AI expertise," might have felt a little less intrusive than asking users to write out personal details, and requests. But alas, this is how OpenAI has decided to roll out customization, so protect your data, and prepare yourself for ChatGPT's custom instruction era.