Writing with language models, without the slop
Prompts as constraints rather than crutches, and where the technology genuinely helps.
26 articles
The failure mode is not that the output is wrong. It is that it is plausible, uniform, and completely without a position.
Where it helps
- Arguing against a draft you are too close to
- Finding the three sentences doing no work
- Turning a rambling voice note into an outline
All three have something in common: the thinking stays yours and the labour is delegated.
Where it does not
Generating the position. A piece with no author has no reason to exist, and readers detect the absence long before they can name it.
A process that works for me
I write the argument badly, in full, in my own voice. Then I ask for the strongest objection to it. Then I rewrite. The model never touches the prose that ships, and it improves the prose that ships considerably.
Objections rather than improvements
“Make this better” produces something smoother and less specific. “What is the strongest argument against this?” produces something I have to answer, and answering it is where the piece gets its spine.
The uniformity problem
Generated prose has a characteristic shape: balanced clauses, hedged claims, an even distribution of emphasis. It reads as competent and forgettable, because it is averaging over an enormous amount of competent forgettable writing.
Good writing is lopsided. It spends four paragraphs on the thing that matters and one sentence on the thing that does not. Averaging destroys exactly that asymmetry, which is why the output feels wrong even when every sentence is defensible.
The tell is not that a sentence is bad. It is that no sentence is more important than the others.
Where I have changed my mind
I was dismissive about using it for research and I was wrong. Not for facts — for orientation. Asking what the main positions in a debate are, then going and reading the actual sources, is genuinely faster than finding the shape of a field by wandering through it.
The rule I settled on: it can tell me where to look. It cannot tell me what is true, and it certainly cannot tell me what I think.