How to Write with an LLM

Original Article ↗ Hacker News Discussion ↗

Writers are grappling with how – or whether – to use large language models in their process, especially for blog posts, documentation, and technical writing. Many favor using LLMs only as strict copyeditors or fact-checkers while insisting that humans draft and choose every word, warning that AI-shaped prose feels generic, erodes personal voice, and floods the web with low-value “slop.” Others argue that, used carefully, LLMs can speed up research, help non-native speakers, and surface stylistic tics, but even they emphasize skepticism, human judgment, and clear limits on how much AI influences the final text.

Overall reaction to the article

  • Many commenters found the piece surprisingly good and practical, especially the “use LLMs as copyeditors, not ghostwriters” framing.
  • Several liked the emphasis on not accepting model wording, only structural feedback.
  • A minority dismissed the whole idea with “best way to write with an LLM is: don’t,” seeing the headline as misdirection.

LLMs as copyeditors and reviewers

  • Popular uses:
    • Flagging passive voice, run-ons, repeated words (“just,” “very”), and unclear sentences.
    • Role‑playing target readers to stress‑test arguments.
    • Fact‑checking technical posts or code and pointing to documentation, with the caveat that models can be confidently wrong.
  • Advocates stress strict prompts: “no praise, no rewriting, only point out issues,” and making the human do all actual rephrasing.
  • Some argue LLMs are still worse than human proofreaders at judging clarity or audience impact.

“AI slop,” trust, and reading fatigue

  • Strong anxiety that ubiquitous LLM output devalues writing, makes readers suspicious, and reduces motivation to read deeply.
  • Several refuse to read anything labeled as AI‑written or that “smells” like it, calling it a waste of time.
  • Others are more pragmatic: if the information is correct and useful, they don’t care who or what wrote it, especially for manuals or specs.

Voice, style, and contamination

  • Many fear that regular exposure to LLM wording will “taint” a human’s voice and normalize bland, influencer‑like prose.
  • The recurring use of phrases like “load‑bearing” is debated: some see it as an intentional joke about AI tells; others say it still breaks immersion and undermines the article’s own advice.
  • Some think style is overrated and content matters more; others argue that writing is thinking, and outsourcing wording outsources thought.

Use cases in code, docs, and fiction

  • For code: models are widely used to check idiomatic patterns or summarize diffs, but some developers now insist on writing their own commit messages and PR descriptions to ensure understanding.
  • For documentation and manuals, opinions split: some see LLMs as fine for structured, low‑art text; others say unreviewed “runbook slop” is dangerous and immoral.
  • For fiction: brainstorming related works and influences seems helpful; generating plot ideas and twists is often described as flat and uncreative.

Automation, ethics, and skill development

  • Some feel guilty replacing human work with LLM‑assisted pipelines but admit it gives a competitive edge.
  • Others see LLM use as just another form of automation, no different in principle from other tools.
  • Multiple commenters worry that relying on LLMs for writing and research will hinder people from becoming genuine subject‑matter experts or strong writers.