OpenAI gives users a new way of prompting in ChatGPT through custom instructions that will be included in all future chats.
In beta, these will first be offered to paying users on the Plus plan before being made available to all users in the coming weeks. The instructions include information that ChatGPT should consider for each chat, as well as guidance on how ChatGPT should respond.
Users will only need to enter their instructions into ChatGPT once, and the AI chatbot will follow them in all future conversations. OpenAI sees this new feature as a step forward, especially for those who use the model professionally.
Teachers, for example, could now work more efficiently by not having to repeat that they are teaching a third class, or developers by being able to determine their preferred programming language.
But also private users would benefit by not having to enter how many mouths they have to feed in their own family every time they make a shopping list.
Starting today, you can set custom instructions in ChatGPT that will persist from conversation to conversation. 👀 📌
You can enable custom instructions in the beta panel from the settings. pic.twitter.com/ljYy3HDaHt
—Logan.GPT (@OfficialLoganK) July 20, 2023
“One of the things I am most excited about for custom messages is conveying my personal writing style,” said Logan Kilpatrick, OpenAI’s manager of developer relations. “I love having ChatGPT help me write and not having to copy and paste my favorite prompt back and forth is going to be super helpful!”
Plugins become more useful and data is secure, says OpenAI
The new prompt option should also make it easier to use plugins by sharing relevant information with them. For example, a plugin that helps you book a restaurant could remember which city to search in.
OpenAI emphasizes that it respects the privacy of its users and removes information that could identify individuals before processing it to train new models.
You can also exclude your data from training by toggling the appropriate switch in the settings, though it’s not entirely clear how much this limits the use of the individual query.
OpenAI has also adapted its security measures to the new ways users can issue commands to the model. If these violate the usage guidelines, they will not be saved.
Customization is the logical evolution
According to OpenAI, the concept of customizable instructions is a response to feedback from users in 22 countries. They asked for more influence over how conversations with the model are shaped.
A major criticism was the constant redefinition of preferences at the start of each conversation. OpenAI had already demonstrated a solution to this problem with the system prompt when it introduced GPT-4.
Currently, individual prompts are not yet offered to all Plus customers of ChatGPT, except for the EU and UK for the time being. As soon as this is the case, it will have to be enabled in the settings under “Beta Features”, analogous to the code interpreter and the (currently unavailable) browsing mode.