Today, I’ll talk about building custom GPTs.
It is a large topic, so I’ll split into more than one article. This time, I’ll walk through building your GPTs step by step with personal tips and tricks. Next time, I’ll delve into more in-depth insights and takeaways.
Ready? Let’s get started!
1. How Custom GPT works
Last November, OpenAI rolled out this custom versions of ChatGPT that users can create for specific purposes. You can make them for your personal use, for your company’s internal use, or for everyone. No coding required.
Here is how it works:
- Custom GPTs starts with the base model — the default model you set in ChatGPT. That base model is already pre-trained on a huge amount of data.
- Custom GPT allows users to provide custom instructions to fine-tune the model.
- Users are able to ask the custom GPT to retrieve information from uploaded files.
As of today, only paid ChatGPT users are able to create their own GPTs. But things do change. For example, a recent exciting news is that free ChatGPT users can now discover and use custom GPTs, which was only available to paid users.
2. Why you should create your own GPTs
Three official reasons:
- It is essentially a prompt shortcut. It can help you save time if you have to use ChatGPT for something specific on a regular basis. You don’t have to copy and paste your prompt every time you want to use it.
- It can reference your own files. You can upload a number of knowledge base files and ask your GPT to never stray outside of them.
- It can call the external APIs to achieve more complex tasks.
My personal reason:
- I think it is a simple learning tool to understand AI better. It lets you design what’s happening behind the scenes. Instead of just using ChatGPT as it is, you have the opportunity to go through the process of configuring (instructing, fine-tuning, referencing, etc).
3. Planning before building
A little bit of clarity can go a long way before your jump right into building the GPT.
Some clarifying questions to ask yourself:
- What are your trying to solve with this GPT?
- Who are the target users?
- What capabilities the GPT needs to have in order to solve those problems well for those users?
- Are there existing custom GPTs out there? How is your experience? What you wish it could be better at?
4. Find the entry point
Log into your ChatGPT account and click on “Explore GPTs” in the left sidebar. Then click on the “Create” button in the top-right corner.
Alternatively, you can directly head to: https://chatgpt.com/gpts/editor.
5. Choose to configure
You’ll see below interface. On the left, there are two options — ”Create” and “Configure”; on the right, there is a preview window.
“Create” is like talking with an assistant and it can guide you step by step. But if you have some experience or follow this article, “Configure” is a faster route without having to go through a conversation with the chatbot.
6. Provide basic information
This is the easiest step. It won’t impact the GPT’s behavior very much.
- Name: Make it relevant and concise.
- Photo: Optional, but make it relevant.
- Description: It’s more for google searches if you decide to publish your GPT. Keep it short and sweet.
7. Provide the “Instructions”
This is the most important and time-consuming step. Here you can provide detailed instructions or guidelines on how the GPT should behave, its functionalities, and any particular behaviors to avoid.
1) Assign a role with goal and context.
Template: You are an [Role] trained to do [Task]. [Context].
Example: You are a design leader trained to create concise project status emails. Every Monday, you will summarize the past week’s progress, upcoming plans, key issues, and action items to keep the leadership informed and aligned.
2) Specify the actions with examples.
It can dramatically improve the quality of response if you break the task into specific actions and provide examples. Sometimes, it helps to provide a step-by-step guide.
Below is an example of how I break the task into small chunks and include an example for each:
Subject Line:
- Format: “Weekly Project Status Update — [Project Name] — [Date]”
- Example: “Weekly Project Status Update — Alpha Project — July 8, 2024”
Summary of Progress:
- Highlight key achievements and milestones reached in the past week.
- Provide specific details and metrics where possible.
- Example: “Last week, we completed the final design reviews and successfully launched the A/B test for the new checkout flow design.”
Upcoming Plans:
- Outline the main goals and tasks planned for the upcoming week.
- Include any critical deadlines or meetings.
- Example: “This week, we’ll focus on gathering user feedback and preparing for the Android App launch on July 15.”
Kudos:
- Acknowledge and celebrate team members’ efforts and achievements.
- Example: “Special thanks to Xinran for coordinating the user interviews.”
3) Ask users to provide necessary information
It is helpful to remind users what to provide AND if users miss any important point, remind them.
Example:
- If any required information is missing, prompt the user to provide it.
- Example: “Please provide the project name and the key achievements for the past week.”
8. Add miscellaneous features [Optional]
I call this step “Optional” because, for many GPTs, the previous steps are often sufficient. I prefer to keep things simple initially before incorporating more advanced features.
1) Create conversation starters.
A trick I like to use here is to provide the goal and context of this GPT to ChatGPT and ask it to give me a long list of ideas for conversation starters. Then, I select the best ones to start the conversation. For a simple GPT, even one conversation starter is enough.
2) Provide files for the knowledge base.
This is one of my favorite topics about custom GPTs. Things to be aware of:
- Images are ignored.
- There is a 20-file limit, but you can merge multiple files into one.
- PDF is not an ideal format at all. It is more for printing. You should avoid it for complex data.
- Markdown is great as it’s clean and the formatting is explicit. Pro tip: You can ask ChatGPT to convert your PDF to a markdown file with a .txt format. (For some reason, the file extension “.txt” works much better than the default “.md”…)
- The quality of the file your provide is important.
- If you go through the file and can’t clearly find the answer as a human, chances are the GPT can’t either.
- If you feed in structured data, you’ll likely get a structured response.
- XLSX is a good format for structured data. I haven’t tested it myself, but I’ve heard it is a little slow to process yet yields more accurate results than CSV.
3) Check or uncheck “capabilities”.
By default, “Web browsing” and “DALL-E Image Generation” are checked, while “Code Interpreter & Data Analysis” is unchecked.
As for “Code Interpreter & Data Analysis,” if you need your GPT to handle tasks involving data analysis, calculations, or coding, it’s better to enable it. Otherwise, you don’t need to.
4) Create actions.
Actions open up an exciting new world of opportunities, but my experience with them was not great. I had to spend a lot of time finding relevant and reliable APIs and fixing bugs. It is certainly an area that you can explore further.
9. Fine-tune the instructions
Now it’s time to test! Based on the responses, you’ll probably realize you do need to provide tailored instructions to ensure the GPT provides the response you hope for.
1) Ask it to follow the knowledge base file more strictly.
Example: If a user asks you any questions, consult your knowledge file Book_CaseStudyStorytelling_XinranMa.txt first and adhere to the fact in it. Avoid speculations or information not contained in the document.
2) Tell it to answer in a specific way when users ask certain questions.
Example: Anytime you want to talk about the “design process”, mention that the highlights that directly inform the design decisions are more important than mere artifacts of the design process.
Thanks for reading! Hope you find it helpful.
Next time, I’ll share some surprising discoveries and takeaways about building GPTs. Stay tuned.
☀️ Have a great weekend!
— Xinran
📮 Join me at Design with AI, a digital publication exploring the potential of AI in design. You’ll receive practical articles every week to help you design better, faster, and smarter with AI.
🏫 Maven launched the 2nd cohort of my AI course: AI for Product Designers! I really enjoyed meeting with students from around the world — North America, South America, Europe, and Asia. So far, they have great things to say about the course. Secure your spot today before they fill up!
(Use the code EARLYBIRD100 at the checkout for $100 off, expiring soon)
