Creating a custom GPT to help with content design
I created a custom version of ChatGPT to see if it could help me as a content designer. In this blog post, I’ll talk about how I set it up, the knowledge I fed it, and my initial thoughts on its performance.
Using ChatGPT for content design
ChatGPT caused a huge buzz when it was made available to the public. Its ability to quickly generate content based on prompts was unlike anything we’d seen before. It gave content designers a powerful new tool to use when researching, writing and editing our content.
I’ve spent a lot of time experimenting with ChatGPT to see if it can help with common content design tasks. There’s undoubtedly a lot of potential, but I’ve found it very limited because it:
- doesn’t use predefined guidelines that set the basic standard for content design
- always responds in American English
- cannot follow content style guides
- needs a lot of follow-on prompts to get the right tone
These issues meant I had to write complex prompts to get close to something that would help me. I spent lots of time copying the same instructions into prompts. For example, I had to constantly tell it to write in British English. Overall, this wasn’t saving me any time and didn’t add a lot of value to my work.
How a custom GPT could help
A custom GPT is a personalised version of ChatGPT, taking extra knowledge, rules and preferences to give you better responses. As a result, you don’t have to start from scratch in every prompt with ChatGPT. For me, this felt like a huge step forward in making ChatGPT a more helpful content design assistant for me.
You can learn more about custom GPTs in this article introducing GPTs from OpenAI.
Creating the custom GPT
My goal was to give it basic content design knowledge and a set of rules that would automatically be followed in every conversation.
Setting up a GPT is done through a chat interface. This is where you tell the GPT how you want it to act. You can change the name of the GPT, give it a custom description and choose an AI-generated graphic. I didn’t change anything that the GPT suggested.
For this experiment, I asked my GPT to follow best-practice content design guidance from sources such as the GOV.UK service manual, Content Design London and Scope.
I also provided more detailed instructions such as:
- use simple positive contractions
- never use negative contractions
- sentences should always be fewer than 26 words
- always write in British English
- never use ‘please’
First impressions
Straight away, I saw an improvement in the quality of the responses. The simple rules I provided were always followed. All answers were in British English, it avoided negative contractions and it never wrote a sentence with more than 26 words. Promising!
Here’s an example of content from ChatGPT:
And here’s the same example from an early version of my custom GPT:
I found it difficult to keep track of the instructions I had fed it. Unfortunately, there isn’t an easy way to see what instructions it’s using. There’s a little box on the ‘Configure’ page, but it’s not clear and doesn’t go into a lot of detail. I started using a spreadsheet to track the instructions I had fed it.
Overall, I was encouraged by my first interactions with my GPT. Almost straight away I fixed some of my frustrations with ChatGPT. It feels like a good starting point but it was clear that this needs ongoing training.
What’s next
Over the next couple of weeks, I’ll test my custom GPT on a range of hypothetical content design tasks. I’ll continue to refine and train the model to give me more useful results. I’m also going to explore how I can better train the model to follow a content style guide.
In my next post, I’ll talk about how this goes and share some examples of its output. I’ll also share a public link to the GPT so you can try it out yourself.