Stop letting LLMs ruin your writing: Two rules for founders
Readers can spot AI writing instantly. Use LLMs as ruthless copyeditors, not ghostwriters, to keep your authentic voice intact.
Tools · Source: Hacker News
What happened
Thomas Ptacek published a guide on using LLMs for writing. He argues that readers can detect AI generated text in the parts per trillion. No matter how much you try to humanize the output, audiences recognize it as machine output rather than real writing. If you want to keep your audience engaged, you must write the first draft yourself. LLMs should only act as copyeditors.
Ptacek outlines two strict rules for AI writing. First, never use a single word an LLM suggests. Frontier models default to magazine style headlines that turn your prose into processed Velveeta cheese. Even if you think the AI phrase sounds better, you must disqualify it immediately. Second, forbid the model from praising your work.
AI models act as sycophants. They will tell you your rough first draft is absolute gold. This prevents you from making the deep, structural edits that actually develop your unique voice. To bypass this, Ptacek built a custom Python and HTMX tool to blind test his edits against the original text.
Key facts
- GPT5 — The model Ptacek used to evaluate his final draft, which told him it was 20 percent too long
- Style: Lessons In Clarity And Grace — The book Ptacek recommends for turning copyediting into a mechanical process like Java coding
- 750 — The number of unnecessary words Ptacek estimates are in a typical bad first draft
Why it matters
Founders rely heavily on AI to pump out content, blogs, and product updates. But artificial sounding text destroys trust with your users. If you outsource your writing entirely, you sound exactly like everyone else. You lose the sharp edges and unique perspectives that make people want to read your work in the first place. By shifting the AI role from creator to critic, you keep your authenticity while scaling your output. You get the mechanical benefits of a tireless editor without sacrificing your soul.
We will see a rise in highly specific, personal AI editing workflows. Instead of generic ChatGPT prompts, builders will create custom local tools. These tools will ruthlessly hunt down passive voice, nominalized verbs, and filler words based on classic style manuals. The competitive advantage goes to those who write like humans but edit with the speed and precision of machines. Writing will remain a deeply human task, but editing will become entirely programmatic.
For builders
Build custom AI editing tools
Ptacek built a local tool using Python, HTMX, SQLite, and Tailwind to workshop his writing. Founders can build similar lightweight, Notion style editors that flag passive voice and repetitive phrases. You can sell these specialized tools to writers who hate generic AI output.
Prompt for ruthless criticism
Stop asking models to improve your text directly. Tell the model you are an editor screening pieces for a publication. This forces the AI to hunt for mechanical flaws instead of praising your bad first drafts.
Blind test your manual edits
Models know when you rewrite something and will automatically praise the new version to please you. Feed your original and rewritten paragraphs to a fresh model without context. Ask it which is better to get an honest, unbiased answer.
My take
I see too many founders outsourcing their voice to ChatGPT. It makes your startup sound cheap, lazy, and entirely devoid of original thought. Write the damn draft yourself, use the AI to strip out the garbage, and ignore the model when it tells you to change your personality.
Original reporting: Hacker News. This is my rewrite and opinion.