How to Write with an LLM

Who Should Read

A developer whose READMEs, blog posts, or PR descriptions went flat after LLM polishing

Core Mechanics

• Use the LLM as copyeditor, not ghostwriter
  • Write the piece yourself first, then have a good model find flaws.
  • Readers register LLM-written paragraphs as output, not writing.
• Rule 1: use no word an LLM suggests
  • Frontier models are great at pleasing phrasing, which makes everything read like magazine headlines.
  • Even if you like a phrase or are sure it's better, don't use it.
• Rule 2: forbid encouragement and distrust praise
  • The LLM praises structure, paragraphs, and metaphors in your draft.
  • That makes you double down on first-draft impulses instead of rethinking.
  • Forbid encouragement in the prompt, and stay vigilant about praise anyway.
• Models are good at spotting problems
  • They flag passive voice, nominalized verbs, repeated phrases, and filler like 'very' and 'actually' without tiring.
  • They also find paragraphs that would be clearer elsewhere.

Community Reactions

• Experience writing PR descriptions by hand
  • One commenter writes commit messages and PR descriptions themselves and only has an agent fact-check them.
  • They say it deepened their understanding of agent-generated code, which is otherwise easy to skim.
• Fact-checking experience
  • A technical blogger said a fact-check prompt pointed to documentation that undermined a whole post and led to a better one.
  • Another commenter said LLMs have terrible taste in style feedback.
• Opinion that the advice is circular
  • Judging LLM suggestions requires already having taste and writing knowledge.
  • They suggest letting a draft rest and rereading it as an alternative.
• Experience asking for specifics
  • Asking it to flag passive voice and run-ons worked well.
  • Without demanding specifics, it gives vague advice and repeats 'trim 10-15%' on any input.

How to Apply

• Ask the model only to flag problems
  • Paste today's README or PR description and ask for passive voice, repetition, and filler with locations.
  • Forbid rewriting.
• Add a no-encouragement rule to prompts
  • Start every copyedit prompt with 'no praise or encouragement, only specific problems'.
  • Rewrite flagged parts yourself instead of copying suggestions.
• Compare rewrites in a fresh session
  • Paste the original and your revision into a new chat in shuffled order and ask which is better.
  • Don't say which one is the revision.
• Build a list of editing prompts and run passes
  • Keep one prompt per issue such as passive voice, nominalization, and repetition in a file.
  • Run each once over a piece.

Try It

# Paste into Claude or ChatGPT (the article's method turned into a prompt)

You are a strict copyeditor. Do not rewrite the text below; only find problems.
Rules:
- No praise or encouragement. Only problems.
- Do not suggest replacement sentences or words.
- For each problem, quote the original and give the reason.
- Look for: heavy passive voice, verbs turned into nouns, repeated phrases, filler like very/actually/really, paragraphs that would be clearer elsewhere.

Text:
{{document}}

Next: put your own rewritten paragraph and the original in a new chat in shuffled order and ask only "which is clearer?"

Terminology

Related Papers

• AI is a tool and using it well takes skill

  A rebuttal to Hickey: LLMs are tools, and skill in directing and checking them decides the outcome

• Reflective Prompt Optimization of a Deep-Research Agent: A Topic-Disjoint, Multi-Seed Protocol for Measuring Transferable Gains and Prompt-Level Overfitting

  Reflective prompt optimization raised held-out citation reward from 0.702 to 0.781 on a deep-research agent.

• GPT‑6 and Intelligent UI for everyone

  GPT-6 reportedly adds generated UI to chat; with no article text, only comment reactions are available.

• Base Models Can Reason By Taking a Cue From Training Data

  Prefilling just two opening tokens lifted Olmo-3-7B's MATH-500 accuracy from 42% to 78%.

• Labels Override Definitions in Jev-Style Typed Decision Models

  Renaming options to A/B raised routing accuracy by +0.1511 over meaningful labels.

• How Much Prompt Is Enough? A Blackbox Minimization of Few-Shots in LLMs

  Delta debugging trimmed few-shot prompts by 65.3% on average with unchanged propositional output

Related Resources

• https://sockpuppet.org/blog/2026/09/17/how-to-write-with-an-llm/
• https://lambdaland.org/posts/2026-08-07-ai-writing-stupid/
• https://writing.brlewis.com/posts/2026/llm-copyeditor/