Here is the strange thing: the tool everyone assumes is built for generating text is actually better at improving it.
Most people use language models backwards. They open a blank document, describe what they want, and wait. What comes back is fluent, well-structured, sometimes impressive in the way a generic hotel room is impressive — everything in the right place, nothing that belongs to anyone in particular. The prose is competent because it has averaged across an enormous range of competent prose. It sounds like writing. It doesn’t sound like you, and more importantly, it often doesn’t say anything you actually think. It says what a reasonable person might think, in a reasonable way, and that is exactly the problem.
Contrast that with what happens when you arrive with something rough. Say you’ve spent twenty minutes getting your actual thinking onto the page — choppy sentences, a tangent you’re not sure belongs, a phrase you can’t quite make work yet. You hand that to the model and ask it to improve it. The result is different in kind, not just degree. It works with what’s there. Your odd sentence rhythms, the ones that reflect how you actually move through an argument, tend to survive. Your specific examples — the ones only you would have reached for — stay in place. The model seems to understand that its job is clarification, not replacement, and so it clears away the noise while leaving the signal intact.
The reason for this asymmetry isn’t mysterious once you look at it. When writing from scratch, the model has no constraint on voice, no anchor in your particular way of thinking, nothing to push against. It defaults to the center of the distribution, which is where most text lives and where nothing interesting lives. When editing, it has a strong prior: the material in front of it. It can see the grain of the wood and cut with it rather than against it. The idiosyncrasies that make your writing yours stop being problems to solve and become features to preserve.
What this implies for how to actually use these tools is simple and a little uncomfortable: write badly first. The first draft is for externalising your thinking, not for producing something readable. Get the ideas onto the page in whatever form they arrive — incomplete, recursive, awkward. Don’t edit as you go. Don’t reach for the model while the document is still blank. The roughness isn’t a problem to be outsourced; it’s the raw material the model needs to do its best work. A rough draft full of your actual thinking gives the model something real to work with. A blank prompt gives it nothing but latitude, and latitude is where voice goes to die.
The deeper implication is about what these tools are good for, which is making thinking legible rather than generating it. Legibility is genuinely hard. Most people who struggle to write aren’t struggling to think — they’re struggling to surface the thinking they’ve already done in a form that someone else can follow. That gap between having an idea and communicating it clearly is exactly where the model excels. It can see the shape of your argument and help you show it. What it cannot do is have the argument for you.
The workflow, then, is sequential and non-negotiable: think first, write badly second, improve with assistance third. Reversing any step in that sequence produces something that looks like the output but isn’t. Professional, coherent, and empty — the written equivalent of a hotel room that nobody lives in.