I've been using LLMs to help write prose, rather than code.

As discovery tools, LLMs impressed me early on. For me, they've largely replaced Google. And once we have the context loaded, why not hand over the pen? Let the LLM write the report from our joint work. In January 2026 that seemed magical. Six months later, rereading what I'd written, I could see the embarrassing ChatGPT template style throughout.

So I doubled down and started translating the best writing style manuals I know into agentic skills. That improved the agents' writing style. I still had to tell them explicitly to stop leading with conclusions. 😅

Only then did another problem become clear: the voice could sound like mine, yet the argument could still contain non sequiturs. And equally damaging, my opinions get reworded and regressed to the mean. 😱

LLMs generate text by predicting the next token from the preceding context. That context can include a plan, but neither the plan nor the prediction guarantees that the next sentence follows logically from the previous paragraph. In my experience, the sentences can individually sound reasonable while the argument goes nowhere. The writing often feels linear, devoid of high-level wit, and remarkably unremarkable. A one-line joke comes readily; a short story whose setup earns its ending never materialises.

So how can one use LLMs to write prose?

I use them adversarially. I tell the agent to read and give me its impression, leaning hard on the writing style reference books I adopt. Outside a work setting, I can't rely on another human being beside me to review my writing. With the agent, I have a reviewer on tap.

Its suggestions I may adopt; its judgement I give very little weight. Each criticism has to stand on its own merit. A convincing explanation on the screen is no guarantee of sound judgement.

The agent helps me out of writer's freeze. The agent reviews my grammar. But it will edit me out if I let it.