How to train your GPT (Part 1)

OpenAI changed the AI game… again

When OpenAI announced custom no-code GPTs at their event last week, it threw everyone into a frenzy. Many have called it the “App Store” moment for OpenAI, alluding to the release of the GPT Store at the end of the month where builders and creators will be able to publish and monetize their custom GPTs on the ChatGPT platform.

All ChatGPT Plus users received access to build their own GPTs last week, so I spent some time tinkering and experimenting. My first attempt was creating an assistant with the identity of an Italian Sloth Chef to help with meal planning. I called him Chef Luigi Rigatoni (ask him if you should break your spaghetti in half!).

For the Hawaiian pizza lovers, Chef Luigi Rigatoni has provided his expert opinion.

While this experiment was fun, it didn’t really tap into the most powerful features of building a custom GPT. It was just ChatGPT with a set of custom instructions that anyone could copy.

I decided to create something more unique for my second GPT. This time, I wanted to take advantage of:

• Custom Knowledge Base: adding a custom knowledge base of data and documents, rather than relying on the general knowledge of GPT-4
• Capabilities: leveraging ChatGPT’s web browsing, image generation, and code interpreter capabilities
• Actions: connecting the GPT to third-party services to retrieve information or take actions outside ChatGPT

Creating the Year 2049 GPT

I’ve been writing Year 2049 for over 2 years, and I’ve published over 75+ articles on AI and other emerging technologies. So it felt like the perfect use case to build a custom GPT because it can benefit everyone in different ways:

• Existing readers: As I write more articles, it becomes harder for existing readers to find specific insights or snippets I’ve shared in the past.
• New readers: There are lots of AI newsletters out there at the moment. Creating this GPT would help new readers get a taste of the content I usually share before putting their email down, and hopefully convince them to subscribe.
• Myself: I have a hard time remembering or finding what I’ve written about before, so this would help me retrieve, reference, and repurpose my previously published content.

The results were pretty impressive. Year 2049 GPT could answer a variety of questions such as:

• What are some different ways I can use ChatGPT?
• How is AI affecting healthcare?
• What AI tool can help me translate a video?

All the answers it gave me came directly from the Year 2049 archive, with specific references to articles I’ve written. For some questions, it provided answers by retrieving relevant pieces of information from different articles at once.

Try it out yourself 👇

Chat with Year 2049 GPT

The path to get there was not the most straightforward, so I’ll share my process and mistakes so you can do this in 30 minutes, and not waste two days like I did.

This is by no means a perfect guide. GPTs have only been out for a week and I’m still trying to figure this all out. Expect follow-ups and extensions to this guide as I keep building and experimenting.

How to train your GPT

Lesson #1: Give it an action-based identity

To start, your GPT needs an identity.

What is this GPT’s purpose? What does it do exactly? What types of questions does it answer?

And here’s where I made my first mistake 🚩.

My goal for Year 2049 GPT was to create an on-demand AI educator that anyone can access, powered by the content I’ve been writing for 2+ years.

Naturally, this was the first identity I gave my GPT:

This GPT is a friendly AI educator who will answer questions based on the Year 2049 newsletter archive at: https://year2049.substack.com

My GPT started drifting.

Instead of answering questions from the Year 2049 archive, it relied on GPT-4’s knowledge to answer questions.

It didn’t give specific references or examples that I’ve written about and mentioned things I’ve never talked about. It completely assumed the identity of an AI educator, while completely ignoring my request to answer questions using the newsletter archive.

After lots of trial and error, I discovered that a slight reframing into an action-based identity worked much better:

This GPT will use the web browsing feature to answer questions about AI from the Year 2049 newsletter archive at: https://year2049.substack.com/

It worked like a charm:

It neatly summarized the key points from my ”AI and the future of healthcare” post and linked to it when I asked “How will AI impact healthcare?”.

Reframing the GPT’s identity around using the web browsing functionality to answer questions based on the archive helped it stay focused on the task.

Lesson #2: Start with simple knowledge transfer

My second mistake 🚩 was overcomplicating the knowledge transfer from the Year 2049 archive on Substack to my custom GPT.

I tried:

• Downloading PDFs of all my articles, but they were too large and the formatting was all over the place
• Writing a Python script to scrape my articles and save their contents in a TXT file, but I faced technical difficulties and lost my patience trying to debug the code

The solution to my problem was much simpler and would have saved me hours of agony: enabling the web browsing capability and instructing the GPT to answer questions based on the archive on my public Substack website (year2049.substack.com).

Not only was this a simpler solution, it was easier to maintain. I wouldn't have to download every new article I published and re-upload it to my GPT’s knowledge base. I also didn’t need to worry about re-uploading articles I may have edited or expanded. Since it was using the web browsing feature, it would automatically have access to all the latest content.

Lesson #3: Write helpful conversation starters

I hate when AI assistants just have a blank chat page staring at you. It reminds me of staring at a blank page while trying to write essays growing up.

Your GPT should always give some hint or guide to help people understand what they can ask, because nobody knows it as well as you do.

Thankfully, GPTs have the option to customize conversation starters that appear when someone launches a new chat. These are a great opportunity to highlight your GPT’s abilities.

In my case, I wanted to highlight use cases based on content “categories” I’ve published in the newsletter:

• Case Studies: How will AI impact my industry? (e.g. healthcare, education, etc.)
• Tools: What AI tool can I use to do X? (e.g. generate a logo, translate a video, etc.)
• Education: What are some free resources to improve my AI knowledge?

💡 FYI: The GPT may change these conversation starters without telling you, so I suggest you monitor these regularly and save a copy of them in case you need to readjust them.

Lesson #4: Fail gracefully

I haven’t written anything and everything you need to know about AI, so some questions can’t be answered with the Year 2049 archive alone.

In this case, I’ve decided to make the GPT fail honestly and gracefully:

Instruction:

If this GPT is asked a question that can’t be answered from the Year 2049 newsletter archive, it should honestly say that it doesn’t have an answer for that yet. It should then tell the user that new AI insights, guides, and case studies are sent every Friday in the free Year 2049 newsletter. It should encourage the user to subscribe to get the latest updates at: https://year2049.substack.com/welcome

There’s no shame in not having an answer for everything, and it’s an opportunity to create excitement for my upcoming work and get new readers to subscribe.

One thing I’d like to investigate for Part 2 of this series is the ability to capture these requests that I haven’t written about yet. It would help me make this newsletter much more valuable!

Bonus: Voice & Tone

This is less of a requirement and more of a nice-to-have if you want to differentiate your GPT even more. By default, your GPT will sound like good ol’ ChatGPT.

Why not give it some additional instructions to make it more different? Some ideas for voice and tone:

• Traits: friendly, confident, sassy, sarcastic
• Verbosity: concise, bullet points, detailed paragraphs
• Vocabulary: no jargon, Grade 5 reading level, expert
• Accents: pirate, cowboy

Ultimately, keep your users in mind and make sure your GPT is as useful as possible in helping them achieve a specific goal.

Next Steps

I’m not done building and improving the Year 2049 GPT, and I plan on investigating the following for Part 2:

• Capturing topics of interest or requests submitted by users
• Uploading custom documents to supplement the knowledge base
• Referencing my video content from TikTok and Instagram
• Auto-subscribing new readers directly from the Year 2049 GPT
• Exploring the Actions feature and what useful third-party services I can integrate

Update: Read Part 2