Markdown Is the Best Interface Between Humans and AI

Have you ever given an AI a perfectly reasonable Word document and then spent half the conversation explaining what parts actually matter?

The document might contain the information the AI needs. But it also contains page breaks, formatting, fonts, layout, and a lot of other details that exist primarily to make the document look right to a human.

For humans, that is useful. For AI, much of it is noise.

This is why I increasingly think we should use Markdown files much more often when working with AI. Not because Markdown is fashionable, or because everyone needs to become a programmer. It is simply a very good middle ground between human-readable documents and machine-readable structure.

The structure is the point

Markdown is, at its core, a plain-text format for structured documents. A heading is represented by a #, a list by a dash, a code block by backticks. The structure is visible directly in the text rather than being hidden inside a document format.

Consider a simple workflow description:

## Customer onboarding

### Inputs

- Customer name
- Contract
- Start date

### Steps

1. Create the customer account
2. Send the welcome email
3. Schedule the kickoff meeting

### Result

The customer is fully onboarded and ready for delivery.

There is not much to parse.

But there is also not much ambiguity.

The heading tells the AI what follows. The distinction between inputs, steps and result is explicit. Lists tell it what belongs together. The sequence tells it what comes first.

This is important because good instructions depend on structure. An AI does not just need the right words. It needs to understand how those words relate to each other.

Markdown gives us a simple way of expressing those relationships.

AI already knows a lot of Markdown

There is another reason this format makes sense.

Markdown is not some obscure notation that we have invented specifically for AI. It has been used for years across technical documentation, websites, repositories, wikis and other text-heavy environments. Large language-model training corpora also contain substantial amounts of material originating from software repositories and other structured text sources. For example, the Pile, a widely studied dataset for language-model training, included a large GitHub component alongside sources such as Wikipedia, ArXiv and Stack Exchange.

Less document, more information

Rich document formats are optimised for presentation. Markdown is much closer to information itself.

This distinction becomes particularly useful when a document is primarily there to be processed rather than printed. A Word document can tell you that a paragraph is 11pt Calibri, has 6pt spacing after it, and starts halfway down a page. Markdown simply gives you the paragraph.

Lightweight by design

There is also a more practical advantage: Markdown files are tiny.

They are plain text. That makes them easy to store, search, copy, transmit and process. It also means that the path from file to AI is usually straightforward. A Markdown file can be read as text and its structure is already visible in that text. There is less work involved in extracting meaning from the container file before the AI can start working with the content.

The same file works for humans

This is where Markdown becomes particularly interesting.

A format designed only for machines would not be very useful for collaboration. Markdown is different because the source remains readable to a human. You can open a .md file in a text editor and understand it immediately. Add headings and lists, and the structure is apparent even without rendering it. At the same time, the same file can be turned into a polished human-readable document.

Tools such as Pandoc can convert Markdown into HTML, PDF, Word documents and many other formats. Its underlying model is specifically built around parsing a document's structure and converting it between representations.

This gives us an important separation:

The working format does not have to be the final format.

We can collaborate with AI in Markdown and publish the result somewhere else.

This is not just for programmers

Markdown has strong roots in software development, and programmers have been using files like README.md for years. That history is part of why the surrounding infrastructure is so mature.

And thats nice because we do not need a special AI application to understand the concept. The infrastructure already exists.

Markdown as a shared working layer

This is probably the most important shift in thinking. We often treat documents as something that humans create and AI consumes. I think we should start treating some documents as a shared working layer between humans and AI. The Markdown file becomes the thing both sides manipulate.

We add context.

AI reorganises it.

We refine the instructions.

AI executes the workflow.

The result can then be rendered into whatever format a human ultimately needs.

Where do we go with this?

I am not arguing that everything should become Markdown.

Spreadsheets are better for many tables. Databases are better for structured records. Images are better for visual information. Rich documents still make sense when visual presentation is the actual goal. But when the main purpose of a document is to communicate information and structure to both a human and an AI, Markdown is difficult to beat.

It is simple enough for us to read, structured enough for AI to reason about, lightweight enough to move around, and established enough that we do not need to build an entirely new ecosystem around it.

We have spent a lot of time thinking about how to write better prompts. Maybe we should also think about where those prompts live. A plain .md file might be a better starting point than the chat box.