Markdown

Markdown is a way of marking up plain text with simple characters: # for a heading, * for a list item, ** for bold. It was designed so that people would not notice the markup. Twenty years later it turned out to suit machines just as well: chatbots answer in Markdown, instructions for agents are written in it, and websites serve pages to AI in it.

Markup you can ignore

The idea is older than Markdown. On the first wiki, Ward Cunningham’s WikiWikiWeb (1995), text was marked up with ordinary characters: two apostrophes for italics, three for bold, an asterisk for a list item.¹² Wikipedia is still written the same way.³ ¶ ct5bkvzry

In 2004 John Gruber turned this idea into a format for bloggers: Markdown, an “easy-to-read, easy-to-write plain text format”.⁴ He modeled it on the way people already format their emails. The main rule: marked-up text should read like ordinary text, “without looking like it’s been marked up with tags”.⁵ ¶ cz23wsckm

The format spread faster than anyone could pin it down, and every service grew its own flavor. In 2014 Jeff Atwood and co-authors merged them into a common specification, CommonMark;⁶⁷ in 2017 GitHub, where it is used to write project descriptions, built its own Markdown on top of it.⁸ That is how the bloggers’ format became the language of programmers. ¶ c2vh2z2ac

When the reader is a model

A language model needs the same thing a person does: to see the structure of a text without extra noise. An HTML page is almost entirely service markup, and a model pays for every token. By Cloudflare’s count, the same post costs 16,180 tokens in HTML and 3,150 in Markdown, 80% less.⁹ This is part of why Markdown became the format in which AI and people exchange text in both directions. ¶ c255uyrek

AI to human. OpenAI states it plainly in its rules for model behavior: unless otherwise specified, responses are formatted in Markdown.¹⁰ Hence the headings, lists and tables in chatbot answers.

Human to agent. For coding agents, project rules go into a Markdown file at the root of the repository. The standard for this became AGENTS.md, “a README for agents”, which OpenAI released in 2025; by December more than 60,000 open-source projects used it, and it is now governed by the Agentic AI Foundation.¹¹¹² ¶ c3qykfvur

Website to agent. In 2024 Jeremy Howard proposed putting an llms.txt file on websites: a table of contents for models in Markdown, because it is “the most widely and easily understood format for language models”.¹³ In 2026 Cloudflare began serving agents Markdown instead of HTML on request, calling it “the lingua franca for agents”.⁹ ¶ cp573dt8e

Wikipedia: same idea, different path

Wikipedia is older than Markdown and is written in its own wiki markup, of the same kind as WikiWikiWeb’s.³ But it is not served to machines in Markdown: in 2025 Wikimedia Enterprise released articles for machine learning in JSON, “instead of scraping or parsing raw article text”.¹⁴ Converting Wikipedia to Markdown for models is so far left to enthusiasts: Santhosh Thottingal, an engineer at the Wikimedia Foundation, wrote such a program as a personal project, because otherwise the service markup has to be stripped out of HTML for models.¹⁵ So the largest wiki, heir to the idea of simple markup, speaks a different language with AI: the language of structured data (see Wikipedia after AI). ¶ c5mwp4xkn

Where Markdown gets in the way

Markup is not neutral. In a 2024 study, the quality of GPT-3.5-turbo’s code translation varied by up to 40% depending on whether the prompt was given as plain text, Markdown, JSON or YAML.¹⁶ ¶ cdmijwrm9

And when a model’s answer is read by a program rather than a person, asterisks and hash signs only get in the way: since late 2024, OpenAI’s reasoning models in the API answer without Markdown by default, and developers turn it on with a special string, “Formatting re-enabled”.¹⁷ ¶ crzeifuan