Markdown won. The review layer didn't come with it.
While playing with the idea of starting Stet, i ran a research project with people who write with AI every week.
The numbers first, so you can weigh what follows. 351 people started. 300 cleared the bar of using AI weekly for work involving writing or documents. 52 finished. Every one of those 52 had edited a markdown file in the past month. Small sample, paid panel, all of it self-reported.
Markdown won
In April, Andrej Karpathy published a gist called LLM Wiki: a pattern for having an agent build and maintain a personal knowledge base. It picked up thousands of stars and forks. Google formalized a vendor-neutral version, the Open Knowledge Format, within days. Garry Tan arrived at the same place from a different direction.
Three people who rarely agree on tooling, all reaching for the same thing: a folder of markdown files versioned in git.
That is not a coincidence and it is not fashion. Plain files, no lock-in, and an agent can read and edit them directly instead of you copy-pasting into a chat window all day. Version history and branching come free. If you moved your documents to markdown in the last year, you were early to the same conclusion.
The review layer did not come with it
I gave the 52 participants four problems and asked them to pick the one closest to their own work. Eleven said none of them fit.
| Problem | Picked it |
|---|---|
| Claude rewrites a paragraph and quietly drops something you needed | 15 |
| Your document moves through two or three AI tools and loses the trail | 13 |
| None of these describe my work | 11 |
| Claude wrote half the spec, you wrote the other half, and later nobody can tell which | 9 |
| You switched to markdown and the review part did not come along | 4 |
Only four picked the version that named markdown and its missing review layer directly. But 37 picked some flavor of the same underlying failure.
A constraint disappears during a rewrite because there is no before-and-after review. A document crosses three tools and nobody can reconstruct what happened because the history broke at every handoff. A line ends up in a spec and nobody knows whether it came from the model, the author, or an edit they ever approved. The surface problem changes. The missing layer does not.
These are not identical complaints, and 37 people did not ask for a markdown review tool. But all 37 chose a problem that review, attribution, or decision history is meant to solve. The file kept the latest text. It did not keep a trustworthy account of how the text got there.
The explicit markdown version came last because it named the category. The other three named the moment someone discovers the category is missing.
Here is the symptom up close, from one respondent:
This actually happened last month when I was using VS Code's AI assistant to rewrite a project proposal in Markdown, and because we couldn't track changes directly in the file, my teammate accepted a major structural overwrite without realizing we lost two paragraphs of crucial technical specifications.
That respondent rated the fit Perfectly and is a writer and editor by trade.
That is the shape of it. In Google Docs and Word, review is part of the writing surface. Comments. Track changes. Suggesting mode. The little back-and-forth with a teammate in the margin. Markdown gives you a portable file, and git can preserve snapshots and diffs. Neither one, by itself, gives a writer a review step around every AI edit or keeps the reasoning attached to the words. The file is current. The story of what changed, who changed it, and what anyone approved lives in Slack, in your head, or nowhere.
Plain text should not mean no review
There is an assumption underneath all of this: that review is a feature of fancy document formats. That if you want suggesting mode and comments, you pay for them in lock-in and a proprietary file.
It is not true. Review is not a property of the file format. It is a property of the tool you open the file with.
Stet opens your plain markdown files and puts the review layer back on top. An AI edit, or a teammate's pushed change, arrives as a proposal. You see what changed. You comment by writing a note right in the document, and your agent picks it up. You accept it, reject it, or send it back.
The files stay plain text on your machine. No new format, no API key, no migration. The thing you lost when you left Word comes back, without the thing you left Word to escape.