The premise of “plaintext productivity” is simple: most software today is overengineered, and the real secret to productivity is to return to radically simpler tools. Once you switch from whatever SaaS-based todo app you’re currently using to a single long todo.txt file, you’ll avoid a million distractions and actually get shit done.
I’ll admit that this concept always seemed slightly performative to me. The enthusiasm from people promoting it often seemed more rooted in ideology than actual productivity (“screw big tech, own your data!”)
But recently I’ve changed my mind and become a religious promoter of plaintext productivity myself. That’s because there’s a new, incredibly compelling reason to adopt this workflow:
LLMs love plaintext files.
This might seem obvious to many, but I don’t think we’ve fully realised all the implications.
CLI tools like Claude Code and OpenCode keep showing people that when their data is organised as simple plaintext files in a logical directory structure, they perform extremely well at almost any task.
Some recent examples:
- Feed raw DNA data into Claude to analyse your health
- Export credit card statements as CSVs to cancel all your subscriptions
- Use Obsidian and Claude Code to build a personal OS
There are a thousand other stories like this all pointing to the same pattern: the “file over app” philosophy is suddenly paying massive dividends in the age of AI agents.
If a workflow is hard to automate today, it’s almost certainly because:
- The state lives on someone else’s computer, or
- it uses an obscure or proprietary file format
Sure, sometimes you can still get the data through authenticated API requests and MCP, but it’s still an order of magnitude more complicated, restrictive, and slower than simply letting an agent go wild with find, grep and ls in a folder with raw text files.
My own computer habits are changing rapidly because of this. I used to be more of a visual person: macOS, polished app UIs, everything had to feel intuitive. Now the visual interface has become secondary. The main thing I care about is that my data is readable by Claude. I write as much as I can in markdown, and spend an increasing amount of time migrating data from all my apps to plaintext formats like CSV, just to see what an agent can do with it.
What happens when everyone starts optimising their software for LLMs over humans?
My hunch is that we’ll all want to make our data a lot simpler, more readable, and more portable.
Turns out the plaintext productivity guys were right all along.