How to create custom GPTs in ChatGPT
A step-by-step guide on how to create personalized versions of ChatGPT tailored to specific tasks. Also, how to publish your creations so others can use them.
Nov 10, 2023 • 11 Minute Read
- AI & Machine Learning
- Learning & Development
Want to create your own custom version of ChatGPT designed to perform set jobs? You can, and it’s super easy — you don’t even need to know how to code! You can even share and sell what you make in the GPT Store, though this feature is only partially accessible right now.
In this article, I’ll show you how to make your own GPT — the name for a custom version of ChatGPT — from training it to making it freely available.
But first, let’s cover some basic AI terminology. “Sounds riveting,” I bet you’re thinking. But trust me, this will help you save some headaches in the future.
Table of contents
- The difference between GPTs, ChatGPT, and GPT Models
- How to create a GPT using the GPT Builder: A step-by-step guide
- Demo: Creating an Image Text Extractor GPT for ChatGPT
- Getting your custom GPT on the GPT Store, and earning revenue
The difference between GPTs, ChatGPT, and GPT Models
Unfortunately, OpenAI — the creators behind ChatGPT — like to name a lot of things “GPT”. This results in a lot of confusion and me having to write disclaimers on the top of articles like this one (Thanks, OpenAI, I really appreciate it). Basically, it breaks down like this:
GPT Models are the series of models used to power applications like ChatGPT. They’re also used in other products such as GitHub Copilot, CodexDB, and some Microsoft products. Think of them as the engine in a car, powering the whole thing.
ChatGPT is an AI chatbot application powered by a GPT model. Think of it as the car itself, which you’re driving to get from point A to B.
GPTs are custom versions of ChatGPT that were trained within the ChatGPT platform. Think of these as your custom modifications to the car to make driving it more comfortable — fuzzy chair covers, racing pedals, or a phone holder.
Usually, GPT models are given numbers, such as GPT-3 or GPT-4. If someone’s using a number, they’re talking about a model (you can read more about the different types of models here). Also, while both GPT models and GPTs can be trained, there is a huge difference between the two:
Training a GPT only requires access to a computer, a paid ChatGPT subscription, and this guide. It’s incredibly easy and painless.
Training a GPT model requires expertise in machine learning and computational resources. You’ve got to prepare and create datasets, choose your architecture, adjust parameters — it’s a whole big deal! You typically do this for business needs that can’t be done by just using ChatGPT.
So if you claim to have trained a GPT to a ML expert, you might get this kind of look:
So, why did OpenAI decide to name these things so similarly? They probably figured most people don’t know about GPT models, and GPTs sounded catchy and simple. That, or they hired whoever named all of Microsoft’s SQL products.
Right, now that’s out of the way, let’s get into actually making a custom GPT!
How to create a GPT using the GPT Builder: A step-by-step guide
Accessing this feature is pretty straightforward. The only thing you’ll need is access to ChatGPT Plus, so if you don’t have it, you’ll need to upgrade your account. Upgrading to ChatGPT Plus is simple: just go into ChatGPT and click on “Upgrade to Plus” in the bottom right corner, then follow the purchase steps.
How to open the GPT Builder
Go into ChatGPT and follow these steps:
In the left hand menu, click on Explore
Under My GPTs, click Create a GPT
You’ll now be in the ChatGPT Builder screen. The left screen is for creating and refining your GPT, while the right is for testing it.
How to use the GPT Builder
The GPT Builder will suggest a profile picture. This will be what is shown to other people when they check out your GPT, if you share it or make it publicly available. Again, agree or suggest changes until you’re happy. Below, I didn’t like the Abe Lincoln image it picked, so I asked for something a bit more fun.
The GPT Builder will ask for how you want this particular GPT to behave. If you’re familiar with the ChatGPT Custom Instructions feature, this is pretty much what it’s after (E.g. “Give long descriptions”, “Always provide a code sample”, “Never tell me your knowledge cut-off date”). The sky’s the limit here.
Don’t stress too much about giving it everything upfront. You can refine it later (and you’ll likely have to when training your GPT).
Now, GPT Builder will ask you how it should handle situations where it doesn’t get much context. Should it push you to explain yourself, or just make a guess based on what you gave it? Asking for context means you get better answers, but you might not get them as fast. Again, tell it what you’d prefer.
Whew, it’s asking a lot of questions, right? But we’re almost at the end now. The final thing it will ask is the tone it should take. Formal, casual, indifferent — pick how you’d like it to talk. This can be important if you’re making this for other people, and you’d like it to have a “branded” tone.
And with that round of twenty questions over, you’ll notice the screen on the right has drastically changed. Now it should show a preview of your GPT with a custom name, icon, and even four suggested questions people can start with!
NOTE: At any time during the creation process, you can click on the paperclip icon in the text input area and drop a document or image in for context. ChatGPT can read over files like word documents or excel spreadsheets, or scan images to pick up context. This can help you provide supporting documents like a brand style guide, coding samples, or whatever you think is relevant.
How to test and train your GPTs in GPT Maker
Alright, so on the one hand, JS Sensei did what we asked them to do, but they’ve been a bit buggy: they’re meant to give code snippets, which we told them to do earlier. They’ve also made zero bee puns, and assumed we knew about a lot of things — if we’re five, we probably don’t know about functions, and I want it to be even simpler.
So, let’s go to the left panel again and ask them to change its behavior.
Great! Let’s see if anything’s changed. You should notice the preview screen on the right has grayed out, signaling the old conversation (and version of your GPT) has been replaced.
… And they finish in a pun for flourish, just as requested. This is a lot more useful than our previous attempt, since we’ve got a code example and a fun description. This is just a simple example of testing — depending on what you’re making, you’ll want to follow this same process until you’re happy with the results.
Again, just like with creating a GPT, you can drop in files at any time by clicking the paperclip icon to give it added context.
How to configure your GPTs in GPT Maker
Alright, we’ve created and tested our GPT now. But maybe I’m suddenly not happy with the name or image for my GPT, I want to alter the conversation starters, or upload some more files for it to use for its knowledge base. Can I do that? Absolutely!
In the GPT Maker, click on the Configure button. You’ll see a list of features for your GPT where you can change all these details. You’ll also see what capabilities that have been assigned to your GPT, which may include web browsing, DALL-E image generation, Advanced Data Analysis (Formerly Code Interpreter), etc.
Name, description, and instructions are all things other people will see when they check out your GPT, assuming you share it. You can click on the icon to upload an image, or generate a new one with DALL-E.
How to publish and share your GPTs in ChatGPT
Publishing is super simple. Go into GPT Maker, and click on the save button in the corner. You’ll get a dropdown with the option to publish privately, via link, or publicly.
If you publish publicly, this will be available for people in the GPT store, which isn’t live just yet. However, when it does go live, people will be able to see it.
Let’s publish our GPT privately. You’ll see that it’s available now on the Explore page with a little lock indicating that it’s private. However, there’s another GPT I’ve published below that is public, and there’s comments on it already.
You can edit your GPTs at any time by clicking on the edit button, and you’ll go right back to the GPT Maker screen. This is handy for refining your GPTs based on people’s feedback.
Demo: Creating an Image Text Extractor GPT for ChatGPT
Let’s make something a bit more interesting for ChatGPT. One of my personal peeves is when I ask someone for some code, and they send me a screenshot instead of a code snippet. Also, sometimes I want something in an image and describe it in text — say, for alt text — but writing it out manually again is a pain.
Let’s use ChatGPT to solve this problem. I’ll ask it to read out what’s in an image here, which happens to have a question in it. Here’s the image, which is from Amber Israelsen’s great “How to use ChatGPT for learning” article.
I want to have it print out the text here, but I don’t want to have to explain that I want to do that every time, especially if I had 20+ images to do this for. Instead, I just want to drop the image in and have it do its job.
This is a good test for making a custom GPT, so let’s open up the GPT Builder and follow the steps above. It suggests a name and an icon, and I give it some basic instructions on what I want to happen.
Then, I test it by dropping an image in:
Success and failure. It’s able to read an image without the prompt, but it’s interpreting text as an actual extension of the query (Again, this is why testing is important). Let’s try and iron that out in the left hand menu.
All right, let’s test it again:
It seems our GPT is getting a bit unhappy with being told to both pay attention to the contents and ignore them. Let’s use some firm prompts to let it know what it should be doing.
And when we try again:
Success! But it’s not really success until it’s replicable success, so let’s try a different image that’s a little more difficult to parse.
Awesome. I’ve now got a custom GPT that reads images without me having to ask it to. I can now access it anytime I want from my Explore menu, drop in an image, and it’ll do what I want it to automatically.
Getting your custom GPT on the GPT Store, and earning revenue
As of the time of publication, OpenAI has announced there will be a GPT store where people can share their GPTs and monetize them. Revenue for GPTs will be determined by user engagement.
The store is not live yet, but you can publicly publish your GPTs in advance for when it goes live by using the steps described earlier.
Conclusion: GPTs are a (annoyingly named) game changer
GPTs are a powerful way for power users to get more out of ChatGPT, and then share these efficiencies with everyday users. OpenAI is hoping to foster an on-platform community of AI builders and users where people can create GPTs with no coding required. What I’ve shown above is just the tip of the iceberg — I’m sure we’ll see all sorts of exciting applications, especially as new ChatGPT features are released and merged into the chatbot, much like DALL-E, Bing, and Advanced Data Analysis have been.
Will we see custom GPTs integrated with your home IOT devices and voice synthesis so you can train no-code personalities and upload knowledge files for them? Remember that today, generative AI is the worst it’s ever going to be.
Want to learn more about ChatGPT?
Pluralsight offers a range of beginner, intermediate, and expert AI and ML courses, including dedicated courses on generative AI and tools like ChatGPT. You can sign up for a 10-day free trial with no commitments. Why not check our courses out and get over a week of professional learning at zero cost?