Notes on writing with AI, reviewing what it changes, and building Stet in public.
Reid Westwood described the document problem AI makes obvious: one decision can cross five files, while every review tool still treats the files separately.
Every tool for reviewing what agents wrote looks like a control panel. Panels, queues, diffs, terminals. That is the right shape for code and the wrong shape for writing, and the difference is not decoration.
Anthropic flipped Claude Code to auto-approve by default. The research behind that call is 80 years old, and every review tool should read it.
You moved your docs to plain text so AI could work in them. Track changes, comments, the back-and-forth got left behind in Word.
A building-in-public look at what an early value-prop study told me. Half of them said the problem was theirs. The framing I was most invested in came last.
If you moved your docs to markdown and the agent suddenly just gets it, that is not a coincidence. A short explainer on why plain text on git became the format AI work runs on.