Mastering the Art of Writing with an LLM

A Final Warning | Meta / Older | September 17, 2026

Writing this guide feels a bit like bragging; it suggests I actually possess a talent for writing. Whether that's true or not is debatable, but I am certain there is a dedicated group of internet critics who believe my prose is abysmal.

Like most people, I am a cocktail of vanity and insecurity, which makes sharing this advice feel awkward. However, the utility of these guidelines outweighs my discomfort. They are simple, logical, and essential for anyone wanting to use AI without stripping the soul out of their work.

The danger is this: no matter how much you try to "humanize" an AI-generated paragraph, a savvy reader won't see it as writing—they will see it as $\text{output}$. To avoid this, you must shift your perspective:

The Golden Rule: Treat the LLM as a copyeditor, never as a ghostwriter.

The Guardrails: Two Essential Rules

To prevent "LLM-creep"—that slide into the uncanny valley where your voice vanishes and the reader tunes out—you must adhere to these two constraints.

Rule $\text{I}$: Zero Adoption of Suggested Phrasing

You are forbidden from using a single word or phrase suggested by the LLM.

Frontier models are unnervingly skilled at picking "pleasing" language. The problem is that they are essentially trapped in "Magazine Headline Mode." While a great headline is effective, a whole article made of headlines is exhausting.

Even if the AI's suggestion seems "better" or more elegant, it is disqualified. You cannot trust yourself to spot every instance where the AI is turning your voice into processed cheese.

Rule $\text{II}$: Ban All Encouragement

Do not let the LLM tell you that your writing is good.

This is a subtle but destructive "influence campaign." When you feed a draft to an AI, it tends to respond with:

"This is gold, Jerry!"

This is the last thing you need to hear. Your first draft is likely a mess: the flow is disjointed, the paragraphs are clunky, and you probably have $\approx 750$ words of pure filler. If the AI praises you, you will double down on your mistakes reinforce your worst impulses.

---

The Practical Application: Mechanical Editing

If you strip away the ghostwriting and the praise, what is left? The LLM's ability to spot mechanical failures. They are superior to humans at identifying:

• Passive Voice: Overuse or inconsistent application.
• Nominalization: Turning strong verbs into clunky nouns.
• Repetition: Using the same word or phrase too often.
• Filler Words: The "sawdust" of prose, such as very, unfortunately, really, and actually.
• Structural Issues: Paragraphs that would be far more effective if moved elsewhere.

The "Prose Schematic"

For those of us with a programmer's mindset, we often crave a technical manual for language—something like C Interfaces and Implementations but for English. That book exists: "Style: Lessons in Clarity and Grace."

It effectively turns the art of copyediting into something as logical as Java coding. I recommend reading it and using its principles to build your AI prompts.

The Workflow

Instead of asking the AI to "fix" your writing, follow this iterative process:

Crucial Note: When comparing two versions of a sentence, use a separate LLM session. If the model knows you just spent ten minutes rewriting a paragraph, it will tell you the new version is better just to please you (Rule II).

Building the Tooling

I eventually grew tired of managing multiple browser tabs and pretending to be a "stern writing coach." I decided to build a custom workshopping tool.

The Tech Stack:

• Backend: Python & SQLite
• Frontend: Tailwind CSS & HTMX (Local build, no CDN)
• Integration: Codex, Claude, or Antigravity CLIs

The Initial Prompt:

"We’re going to build a writing workshopping tool. 
Python, HTMX for interactions, SQLite backend, Tailwind frontend, use a local build not the CDN. 
Do Genius-style sidebar commentary to match highlighted things. 
Multiple documents, track revisions, allow user to flag major revisions. 
Get me this far and then I’ll tell you what I really want."

Final Checklist for the Writer

• Did I write the first draft entirely on my own?
• Did I forbid the LLM from praising my work?
• Did I reject every specific word suggestion from the AI?
• Did I use a fresh session to compare my manual rewrites?
• Is the final voice mine, or the model's?

Whatever system you build will be better than mine, because a personalized system is always superior to a borrowed one. Be vigilant. Don't let the machine trick you into thinking a mediocre first draft is a masterpiece.

Keep your voice. Use the tool to make the process faster and less painful, but never let it drive.

---

I fed this piece to GPT5 a minute ago (I didn't write this...) $\rightarrow$ Wait, that's the joke.