What markdown actually is
Markdown is plain text with a handful of typing conventions. A # before a line makes it a heading. Asterisks around a word make it bold. A dash starts a list item. Square brackets make a link. That is honestly most of it.
John Gruber released it in 2004, with contributions from Aaron Swartz, to solve a blogging-era problem: writing for the web meant writing HTML tags, and HTML tags are miserable to type. The design goal was stated plainly: the raw text should be readable as is, and convertible to a formatted page when needed.
That dual readability, where a human can read it raw and a machine can parse it reliably, sat mostly unremarked for twenty years. Then AI arrived, and the property that had seemed minor turned out to be the whole point.
Why AI converged on the brick
Nobody held a summit and picked markdown as AI’s document format. It won bottom-up, for mechanical reasons.
First: the models were raised on it. Every major model’s training data includes oceans of markdown: every project description on GitHub, developer documentation, forums, and wikis. When a model formats an answer, markdown is its native dialect, the way your mother tongue is the one you count in.
Second: it carries structure without baggage. A Word document that says “Budget” in bold 14-point Calibri wraps that information in layers of invisible machinery: styles, themes, and an entire proprietary file format. The markdown version, ## Budget, carries the same structure in eight visible characters. Structure the machine can’t misread, at a cost the machine barely notices.
Third, and this is the part I care about most, it is auditable by you. The instruction files that steer AI tools, the knowledge bases they draw from: when these are markdown, they are text files you can open, read, and correct with any editor, on any computer, forever. No app, no subscription, no vendor. There is no company that can go bankrupt and take markdown with it.
Try the counterfactual. A Word file from 2004 needs software archaeology to open faithfully today. A markdown file from 2004 opens perfectly in anything, including programs that won’t be written for another decade. One of these is a format for the next twenty years of accumulated knowledge. It isn’t the fancy one.
What this looks like in practice
I should admit the personal stake: nearly everything I produce lives in markdown files. My notes. My project documentation. The instruction files that tell my AI tools how I like to work. And yes, these cups. The piece you are reading was written as a plain markdown file in a folder, and the AI tools that helped me research and polish it read the same file. No exports, no copy-paste, no format conversions.
When your knowledge is in plain text, humans and AI can finally work on the same document, with no translation layer in between. Most people’s AI workflow is copy out of Word, paste into chat, copy out of chat, paste back, reformat, repeat. The plain-text workflow deletes all of it.
And for a reader who has never typed a hash mark: this costs you ten minutes. Heading, bold, list, link: that is the entire vocabulary. You now write the same language the AI answers in, and you’ll notice your prompts improve too, because a structured request is easier for a model to follow than a wall of prose.
What about Word and PowerPoint
Your company’s accumulated knowledge, from strategy decks to process docs and quarterly reports, is overwhelmingly stored in formats built to look right to a human on a screen or printout. PowerPoint, Word, PDF. Formats that are, in exact proportion to their polish, awkward for machines to read reliably. Anyone who has watched an AI shred a PDF’s table into confetti has seen the tax being paid.
I wrote a piece a while back arguing that if a company can’t organize its website, its data isn’t ready for AI either. This is that argument’s sharper sibling: knowledge trapped in presentation formats is knowledge the AI era will quietly route around. The documents that will compound in value are the ones both your successors and their machines can read without ceremony.
I don’t expect the office suite to die. Layout, print, and contracts have their reasons. But the default, the format knowledge is born in? That default is up for grabs for the first time in thirty years, and the AI tools have already voted.
The brick test
There is a pattern here that has become one of my quiet convictions about technology: when a system has to serve both humans and machines, the simple, open, boring thing beats the rich, proprietary, impressive thing. Not always immediately. But eventually, and then permanently. The web runs on plain-text protocols older than most of its users. And the AI wave, for all its exotic hardware, speaks a blogger’s shorthand from 2004.
Next week, I will look at what happens when you start connecting these bricks: how a little app named Obsidian turns a folder of text files into a living map of your own thinking.
Until then, here is the practical question I will leave you with: think about the knowledge you want to still have in twenty years, yours or your company’s. What format is it in right now? Will it open? And when you ask a machine to read it, because you will, will the machine find a document, or a puzzle?