ChatGPT Rolls out New Personalization Feature with Custom Instructions
Set baseline metaprompts to be applied to all conversations
ChatGPT Plus just got a useful upgrade. Subscribers can now add information about their user preferences and persona, which ChatGPT will use in formulating responses. The new custom instructions feature does not change what you can do with ChatGPT. It simply saves you from repeating your preferences every time you enter a prompt. According to OpenAI’s announcement:
We’ve heard your feedback about the friction of starting each ChatGPT conversation afresh. Through our conversations with users across 22 countries, we’ve deepened our understanding of the essential role steerability plays in enabling our models to effectively reflect the diverse contexts and unique needs of each person.
ChatGPT will consider your custom instructions for every conversation going forward. The model will consider the instructions every time it responds, so you won’t have to repeat your preferences or information in every conversation.
For example, a teacher crafting a lesson plan no longer has to repeat that they're teaching 3rd grade science. A developer preferring efficient code in a language that’s not Python – they can say it once, and it's understood.
And it works. For example, I added a request to confine all responses to 80 words or fewer in the second dialogue box in the image above. I tried several prompts with the command activated and de-activated. The system followed my preference for short responses when it was activated. After I deactivated it, ChatGPT went back to longer answers.
Putting Custom Instructions to the Test
Custom instructions responded to several different style and persona inputs, such as maintaining a balanced response that does not favor one argument over another, and writing like a journalist or fantasy fiction author. Of course, you may have been doing this all along. If you add context and preferences to your base prompt request, most chatbot’s based on large language models (LLM) with take that into account when generating the response.
The new feature enables you to set a baseline persona and preference set so you don’t have to repeat the context in every prompt. And ChatGPT Plus appears to ignore some persona and preference requests if you add them to your prompt and they conflict with the custom instructions. In other words, custom instructions are prioritized where there is a contextual conflict right now, and you must deactivate the instructions to use alternative details.
User Defined Metaprompts
There is a lot of talk about prompt engineering and prompt design. These terms are often conflated, and some people use the prompt engineering term for every prompt activity. However, there are different types of prompts, so definitions can be helpful.
Metaprompts are information that the software appends to a user prompt prior to sending the request to the large language model (LLM). They are not included in the user prompt, but the system ensures they are appended as instructions for the LLM.
A common use of metaprompts is to maintain a consistent style of response, which you can think of as a form of alignment. The model provider employs these tools on the user’s behalf. OpenAI’s custom instructions is a way for you to define a metaprompt that is only applied when you use the AI model.
It may be worthwhile to clarify that I am employing a metaprompt definition similar to usage by Microsoft. There is an alternative definition used by a LangChain developer that has decent SEO and suggests metaprompts are a system learning feature. I would argue he is explaining a particular type of metaprompt, not the whole category. Of course, Microsoft muddies the water here because it uses both the term metaprompt and system message interchangeably!
Baby Steps Toward Personalization
Custom instructions are a useful new feature for ChatGPT, but I wouldn’t overestimate its marketplace impact today. This is not as substantial a change as something like plugins. Granted, custom instructions may get more initial usage than plugins because they are easy to use and provide immediately recognizable value.
Of course, one place OpenAI highlights where customer instructions could become particularly useful is for plugins.
Adding instructions can also help improve your experience with plugins by sharing relevant information with the plugins that you use. For example, if you specify the city you live in within your instructions and use a plugin that helps you make restaurant reservations, the model might include your city when it calls the plugin.
OpenAI says that custom instructions are in beta, still have improvements needed for alignment and safety, and are unavailable today in the UK and EU. That suggests custom instructions are available in other markets where ChatGPT operates.
