I’ve been banging the plain text drum for a long time. The argument always seemed obvious to me: text files are portable, versionable, toolable, and permanent. Binary formats are walled gardens.

For me, the gateway drug was LaTeX. My PhD thesis was written as a plain text document - markup, references, figures, all of it in as .tex files. The compiled PDF looked perfect, but more importantly the source was just text. I could version it, search it, script against it. Every paper and large report I’ve written since has followed the same pattern: a single plain text document that links plots, data, and prose together1. The analysis and the writing live in the same place, and you can trace a number from raw data to published figure without leaving the file.

That experience made the benefits visceral rather than theoretical and I wanted others to see this. But it’s a hard thing to transmit.

People reach for Word because Word was the ecosystem used by others. They built spreadsheets with a mouse because that’s what they were taught. The overhead of the plain text approach - learning a markup language, understanding version control, accepting that your document might not have a drop shadow on the heading - was too much friction to accept the benefits.

Then GenAI hit.

LLMs are text in, text out. Every single thing you want an AI to read, process, write, or edit has to be expressible as text. And that, It turns out, is a Trojan horse.

Watch what’s happening. People who would never have considered writing a Python script to generate a report are now doing exactly that - because they can describe what they want in plain English and the model writes the code. The spreadsheet that used to be clicked together is now a pandas dataframe defined in thirty lines. The Word doc is now a markdown template with a generation script. The PowerPoint is now a Marp or Reveal.js file checked into git.

If you have used these tools (Claude Code, Cowork, GitHub Copilot, etc.) for knowledge work - you have seen this first hand

You may not have chosen plain text because you believed in it. You arrived at plain text because it was the path of least resistance when working with AI tools. The destination is the same. We just approached it from the side.

There’s a video I’ve been pointing people at from the No Boilerplate channel - “How to be a great team” - that articulates the structural case for plain text teams better than I usually manage to. Everything in git. Markdown everywhere. Wiki, issues, decisions, all of it in text, all of it versionable.

It’s the Ulysses pact argument: you tie yourself to plain text now so your future self can’t be tempted away by whatever the hot new app is next year.

What’s striking is that AI is making this pact on people’s behalf, whether they intended it or not. If your workflow needs to survive a handoff to an LLM - and increasingly it does - then your workflow needs to be textual. The market is enforcing what advocates couldn’t.

The irony is rich. The technology that many feared would replace programmers is instead teaching everyone else to work like programmers. Version control, reproducible documents, programmatic generation, separation of content from presentation - these aren’t niche values anymore. They’re increasingly just how things get done.

I’m not complaining. I spent years losing the argument and now I’m watching the argument get won for me, sideways, by a language model that doesn’t even know it’s making the case.

The end state - a world where critical information lives in plain text, under version control, readable by humans and machines alike - looks the same as it always did. We’re just arriving at it from an unexpected direction, and with a lot less convincing required.

Footnotes

  1. This paper was the pinnacle - single org file with attached CSVs. Chefs kiss ↩