A couple of weeks ago, I wrote a post about something I’m calling ‘the golden triangle’. This is the combination of three bits of tech that all work together, behind the scenes, to make a custom GPT work more effectively; namely, Markdown, JSON and Python.
Last week, I gave JSON ‘the full treatment’ and spent a bit of time explaining in more detail what it is and how it works. This week, it’s the turn of Markdown. So, today, I wanted to explore a simple question: how does Markdown make GPT responses more structured, readable and reliable?
As with the JSON post, I’ll look at what JSON is and how it works. And, explain the connection between Markdown and Information Mapping; and why this relationship is important for custom GPTs.
HTML: Markdown’s sibling?
I remember the first time I actually saw HTML tags in action, I was pleasantly surprised at how relatively simple they were to understand and use. After all, something called Hyper Text Markup Language hardly sounds like it’s going to be easy or straightforward!
When I first came across the term Markdown, in the context of custom GPTs, I guessed it might be similar to HTML. And it turned out I was right. Markdown also shares HTML’s simplicity and ease of use.
So, what’s it all about?
Markdown explained
If you have used ChatGPT or similar, you will be fully familiar already with the results of Markdown. You have probably noticed that whenever you get a response from your favoured AI tool, its output is generally well-structured and consistently formatted. This is largely (although not entirely) the result of Markdown.
In essence, Markdown is simple framework which ensures that content is presented consistently to users of a given AI tool. It defines a simple content hierarchy. This consists of a maximum of 6 levels of header. It also defines things like a presentation style for things like bulleted lists, tables, links and the like.
What Markdown doesn’t do
It’s important to note that the Markdown itself doesn’t define precise formatting. So, for example, it defines the header levels of a content hierarchy but it doesn’t define the exact font, font size, font style or colour of each header.
That needs to be decided and set at a local level. So, a Markdown defined Header 1 could look different in ChatGPT than it does in Google Gemini, depending on the format chosen by each company. But in both contexts, it will always be the most prominent header in the content hierarchy.
Markdown’s appeal
This makes using Markdown a very appealing prospect. It applies a level of consistency that everyone can subscribe to while still allowing enough freedom for an individual GPT to customise the actual styling of individual elements within the Markdown framework.
If you are creating a custom GPT or series of custom GPTs it also ensures consistency applied to the presentational output across your GPT(s).
A quick intro to Information Mapping
If you’ve read any of my previous posts, you’ll know that I’m a big fan and long-time user of an approach called Information Mapping for structuring, sequencing and presenting content and information.
At heart, Information Mapping is about chunking with relevance. In other words, breaking information down into short manageable chunks, known as Blocks.
The Information Mapping approach also defines a series of Information Types which help to clarify the purpose of a given Block. Using only one type of information per Block ensures you keep its purpose very focused.
Finally, there are some recommended presentation styles for each of the Information Types. These increase clarity and help improve recall of content.
Mapping and Markdown
I’ve noted elsewhere that applying Information Mapping to the structure of any Knowledge Base documents will be beneficial to the ‘ingestion’ of the content in that document.
However, the Markdown framework takes things a step further. It means that you can align its framework with the Information Mapping content hierarchy and presentation styles.
This sets up your GPT’s output so that, broadly speaking, it follows the presentation modes recommended by Information Mapping for each of its six Information Types.
For me, this is a particularly pleasing outcome, as this ensures that any custom GPT created using PerformaGo can adhere to Information Mapping’s best practices for organising and presenting output content.
So, in summary, Markdown is a simple and easy-to-understand framework which helps to apply consistency to the presentation of your GPT’s output.
And it’s going to be design languages like Markdown and JSON which shape the behind-the-scenes PerformaGo framework, helping to make performance support creation feel effortless.
Until next time…
PS If you don’t know me already, I’m Andrew Jackson, co-founder of Pacific Blue Solutions and founder of Pacific Blue AI. I’m using this Diary to document my journey to create a simple, usable app (called PerformaGo) that makes it easy for L&D (or anyone else who teaches something) to create AI-powered performance support GPTs for their learners.
