How to Write With an LLM: Two Rules That Keep the Machine’s Distance

How to Write with an LLM is a strange piece of advice that asks writers to use AI as a tool without letting it take over. The author, whose name is not given, wants LLMs to find flaws in your writing, not to rewrite it. The core argument is simple: LLMs are great at spotting problems, but they are terrible at producing natural, human-readable text. The trick is to keep the machine’s distance.

The Two Rules

The author lays out two strict rules for working with an LLM. The first rule is that you may not use a single word an LLM suggests to you. The reasoning is that models are supernaturally good at picking pleasing turns of phrase, but those phrases often feel unnatural in context. The author describes LLM suggestions as tasting like Velveeta — processed and artificial.

The second rule is to avoid encouragement entirely. When you hand a draft to an LLM, it will reply with enthusiastic praise. That is not what you need to hear. Most first drafts are bad, with incoherent topic flow and hundreds of words that could be cut. The model will encourage you on your overall structure, then on your paragraphs and transitions, then on your word choices and metaphors. All of it is bad, and all of it pushes you toward a finished product that feels engineered rather than written.

Why the Rules Work

The author admits to a long habit of lying to LLMs, telling them he is not the author but an editor screening pieces for inclusion. This helps, but the model usually overshoots, tailoring its responses to the imagined publication’s goals. The solution is to forbid the model from encouraging you and then be hypervigilant about praise.

The author’s own experience supports the rules. He used to open every copyediting prompt with the lie that he was not the author. That helped, but the model still overfit to his imagined publication’s goals. Now he forbids encouragement outright and watches closely for praise.

What the Models Find Well

LLMs are excellent at flagging problems. They catch overused turns of phrase, repeated word choices, and filler words like “very,” “unfortunately,” “really,” and “actually.” They also spot paragraphs that can be moved elsewhere to improve clarity. The author calls these edits “really, actually, very satisfying.”

A Better Workflow

The author recommends reading Style: Lessons in Clarity and Grace, a book about prose that he compares to “C Interfaces and Implementations” for prose. He found out about it from Richard Gabriel and notes that every programmer he knows should have a copy on their desk.

The suggested workflow involves three steps:

1. Ask the model to spot problems in your writing.
2. Rewrite the paragraph, sentence, or section.
3. Present the original and new writing to the model and ask which is better.

The catch is that this approach runs into a variant of Rule Two. The model knows you just rewrote something and wants to tell you the new version is better. The fix is to give the options to a model that does not have the context of your editing process.

A Warning About Overfitting

The piece ties its advice directly to the risk of LLM-creep, the uncanny valley between expression and output, and the danger that readers will lose attention when they detect LLM words in the parts per trillion. The author’s warning is that LLM-creep can knock you into the uncanny valley between expression and output and knock you out of your reader’s attention.

The advice is worth following. The rules are strict, but the payoff is real. An LLM can flag problems you would miss, and rewriting with your own voice remains the only way to produce writing that holds a reader’s attention. The irony is that the most effective way to write with an LLM is to treat it like a tool you barely trust.

Source material: “How to Write with an LLM,” sockpuppet.org.

AI writing · artificial intelligence · creativity · LLM · technology