ChatGPT Custom Instructions: Get ChatGPT to remember you

If you’re working on a project, especially as a developer, ChatGPT can be a real godsend. While it can’t program the whole thing for you, it can do things like write or review code. The problem is, every time you want it to do that, you’ve often got to give it all the context of your project.

And then, for context, you need to copy-paste in large chunks of code. The first time you do this, it’s not so bad… 

…The tenth time, it’s pretty annoying.

Why so much re-introduction? Firstly, ChatGPT doesn’t remember things between chat windows. While ideally it’d be great to just stick to one chat window per request, sometimes that doesn’t work in practice. 

The reason is sometimes ChatGPT can read your brief and go down a rabbit hole of misunderstanding, and even when you try and correct it, it just seems fixated on reading things a certain way. The solution? Spin up a new chat and re-brief it. And sometimes, you might need to do this three or four times — that’s a lot of copy-pasting code!

Even when there’s no misunderstandings, if you want it to help with a different task on the same project — like working on a UI header after talking about a HP tracker — it helps to spin up a new chat to avoid it thinking you’re still talking about the first thing. 

This means you’ve got to brief it all over again. And again. And again.

After a while, working with ChatGPT can feel like you’re dealing with Dora from Finding Nemo (“Just keep swimming”). And because ChatGPT is pretty great as a springboard for everything from ideation to optimization, you can get stuck in a constant cycle of re-briefing and copy-pasting code.

You’re also not just repeating your brief, but normally your preferences. “Please provide code snippets, not just high-level advice!” you might implore when ChatGPT fails to give you anything actionable for the 209th time. 

Thankfully, the creators of ChatGPT, OpenAI, have released a new feature that helps you solve all these problems: Custom Instructions! In this article, I’ll explain what this feature is, how to use it, and share a demo example.

… And just for funsies, we’ll also show you how to get ChatGPT to always answer you as Yoda (even in your code snippets).

The best way to explain ChatGPT’s Custom Instructions feature is like you’re handing the chatbot a written note. The note explains who you are, what you want, and what your preferences are. The Chatbot reads the note before you start a new chat, and takes it into account when giving its answers.

“Right, I’m up to speed. Let’s go!” It says.

So who writes the note in the first place? That would be you! To do that, you’ll first need to turn on the Custom Instructions feature.

To access Custom Instructions, you need to have a ChatGPT Plus subscription. It’s also a Beta feature, so expect bugs.

To turn this feature on, follow the following steps:

• In the bottom left corner, click on your profile name

• Click on Settings & Beta

• In Settings, click on Beta features
• Activate the toggle for Custom Instructions, then exit the window

Now the feature should be turned on! If you click on your profile name again, there should be a new option called Custom Instructions.

Now it’s time to actually write the reminder for our AI Dory. Click on the new Custom Instructions option, and you should see this window.

Bit daunting, right? It’s broken into two sections: Your bio, and preferences on how you want ChatGPT to act.

Enter anything about yourself that you think that ChatGPT might need to know about you. A good place to start is with your role.  For instance, you might want to mention you’re a developer who works exclusively in C#, or you’re working on a PHD in AI ethics. You can even say you're a father with four children (if you're asking it it help you come up with big recipes to feed them!) That said, you don’t need to give it your whole life story, just enough to give it something to work with.

Other things that might be worth mentioning, assuming they’re relevant to how you want to use ChatGPT, are:

• Your current projects and goals

• Subjects you are interested in

• Where you are based

You’ve got a limit of 1500 words, so you can go pretty nuts here. Here’s an example of something I wrote in as a sample, continuing with the theme of game development (but as I said, anything would work, like telling it you're a cloud computing engineer).

Keep in mind with the following prompt you might want to come in and update this if you’re working on multiple projects, but let’s assume we’re just working on one.

As mentioned earlier, if you’re a programmer, one of the biggest frustrations is that when you ask ChatGPT for some code it can often give you general high-level advice and no snippets! You’ve got to chastise it until it does give you one, and even then, it still just loves being general, and giving oh-so-short answers.

Thankfully, this next section is where you can explain exactly what you’d like to see from it. Want longer answers or those code snippets? Want it to bias towards the most efficient solution? Like solutions presented in table format with pros and cons? Tell it all the things you’d like to see from it. 

Oh, and are you sick of this message?

This is where you can tell it to never tell you that again. It might have a memory like a goldfish, but it doesn’t mean you do. Following on from before, here’s an example of an entry.

You might have noticed the checkbox down the bottom that says ‘Enabled for new chats’, on by default. I’d leave this on, since this is pretty much when you’re going to be using it.

Now that you’re done entering everything, hit okay. Everything should be set up!

For review, here are our custom instructions from earlier. Let’s put them to the test by asking an incredibly basic prompt, and see how ChatGPT does. Because we’re dealing with code and dev work, I’m going to use GPT-4 and the new Code Interpreter feature as well. If you haven’t checked out Code Interpreter, seriously, that should be your next port of call — the interpreter is 🔥! We’ve written a whole article on it.

Anyway, here’s the prompt:

So, what’s the answer?

lready, we can see ChatGPT remembers we’re making a fantasy RPG. It also seems to remember that we’ve asked it to think like a designer.

And what about the code snippets? It’s ticked that box too.

It goes on, since I told it to give longer answers, but already we can see it’s taking our preferences and background into account. Our prompt needed to be a fraction of the length of the example at the start of the article!

With this new feature, you can tell it you’d like it to adopt certain personas, world views, or particular tones of voice. Because I apparently have zero fear of reprisal in the future robot uprising, let’s make it speak like Yoda in every interaction.

Here’s the entry in our Custom Instructions settings. Very basic!

I asked ChatGPT for a very simple mechanic to track HP. Here’s how it responded, true to Frank Oz.

I was delighted to see it had taken on board the advice to write comments in Yoda speak, while leaving the actual code thankfully Yoda-free.

And the last bit of this is just gold.

Of course, this sort of feature has other non-Yoda applications (I know, I know, sounds crazy!). You could get it to take on certain philosophical stances, weight arguments in a certain way, and so on. The social implications of that sounds quite big, but in reality, you’re still handing ChatGPT a note — you probably could have asked it to do that before, but now it’ll just remember it between sessions. 

Custom Interactions looks to be a real value-add feature to the ChatGPT toolset, along with Plugins, Code Interpreter, and Browse With Bing. It’ll be nice when we get a toggle that allows ChatGPT to just remember the general context across multiple conversations without writing a note, but for now Custom Interactions offers a lot of hands-on control and a good compromise.

If you’ve got ChatGPT Plus, go check out this new feature and have fun! I double-dog dare you to make it sound like Scooby Doo (or — *gasp* — Scrappy Doo). I’m just going to sit here and wait for my comeuppance as a newly-sentient AI hunts me down twenty years from now, metal hand clasped around a laser rifle, and then says:

“Made me speak like Yoda, you did.”

Want to learn more about getting the most out of generative AI like ChatGPT? If you enjoyed this read, here’s a few more articles to check out:

• How to use ChatGPT’s new “Code Interpreter” feature

• How to make a ChatGPT plugin

• How to use ChatGPT to write code

• The top ChatGPT plugins for developers

• Prompt Engineering: Is it an actual job?

• ChatGPT Vs Bard: Which is better for coding?

• 7 benefits of AIOps for organizations

And of course, my personal favorite: “Pigeon VS AI: Is a wild bird smarter than artificial intelligence?” It’s both academically interesting, and also has cartoons of a bird fighting a robot. Really, what’s not to love? If you’re after more robust learning and to add some different AI feathers to your skillset, I’d recommend checking out these video courses:

• Amber Israelsen’s “ChatGPT and Generative AI: The Big Picture” and “Getting Started on Prompt Engineering with Generative AI.”

• The “Artificial Intelligence: Foundations” learning path at Pluralsight

• “AI and Office” by ClipTraining

• “Artificial Intelligence Essentials: AIOps (Artificial Intelligence for IT Operations)” by Wilvie Anora