A pattern I've noticed across three recent projects: the parts of the work I expected Claude to be useful for — generating prose, summarising, "ideating" — were the parts where it produced the least useful output. The parts I assumed it would handicap me on — polish, restructuring, naming, the small consistency work that makes a piece feel finished — turned out to be where it shone.

Polish is patience. The model has more of it than I do.

This isn't a post about AI replacing writers. It's a post about how a writer's relationship to the work changes when the most boring part of the job becomes optional.

The model is patient in a way I'm not. It will read every paragraph as if it had never read one before. It will notice that I used "platform" four times in three sentences. It will catch the lede that doesn't actually lede. It will offer five alternatives, three of which are bad in instructive ways.

I've stopped asking it to write. I ask it to read.

I've stopped asking it to write. I ask it to read.

The shape of the prompt that works is small:

// the prompt I send before publishing anything
You are reading this draft. Don't rewrite it. Read it.

For each paragraph, tell me:
  • the verb working hardest
  • the phrase you'd cut without asking
  • whether the paragraph repeats the previous one

Don't soften. Don't diplomat. Read.

The first time I did this, I was suspicious. I expected the model to soften my position — to round the edges, smooth the rhetoric, default to the diplomatic move. It didn't. The corrections it offered were almost all sharper than my originals. "Comprehensive guide" became "guide." "It is important to note" became deletion. "There are several reasons" became "Three reasons." Every change tightened the screw.

What I've stopped doing

• Asking the model to draft. The drafts are mine.
• Asking it to "make this better." The instruction is too vague; the model interprets better as smoother, which is usually worse.
• Trusting the model on stance. If I said something with intent, the model's first instinct is to make it more agreeable. That instinct is wrong.

What I do instead

• I write the piece, then I ask the model to find every paragraph that says the same thing as the previous paragraph.
• I ask it to find sentences where the verb is is and propose alternatives where the verb does work.
• I ask it which sentence I could cut. I don't ask if I should — I ask which.

The work is sharper. The position is mine.

There's a version of this argument that lands the wrong way. Real writers don't let AI touch their work — that's the version I want to disagree with. There's also the version that lands too eagerly: Use AI for everything, you'll be 10x faster! — that's the version I find embarrassing.

The truthful version is small. A working writer has a long backlog of polish moves they don't make because they're tired. The model can do those moves at any hour, in any state, without needing to be paid in attention. That's the work it should do.

The thinking is still ours. If you let it.