// HOW-TO · AI WRITING

How to write with an LLM without sounding like one (2026)

How to write with an LLM as a copyeditor, not a ghostwriter: draft it yourself, reject its word choices, and edit for clarity and voice.

Last verified · 2026-09-18 · by Moe Ameen

The best way to write with a large language model is to not let it write. The most durable advice on this — Thomas Ptacek's "How To Write With An LLM," published September 17, 2026 — puts it plainly: use the model as a copyeditor, never a ghostwriter. Readers detect model-written prose, in his phrase, "in the parts per trillion," and no amount of prompting fully hides it, because frontier models default to a polished, headline-y register that is pleasant one sentence at a time and exhausting in bulk. Hand the model your voice and it quietly launders it into that register.

The workflow that actually works flips the usual order. You write the draft; the model finds the mechanical flaws you are too close to see; and you fix them, in your own words. That keeps every idea and every phrase yours while offloading the tedious part — spotting passive voice, filler, repetition, and paragraphs in the wrong order. This is the step-by-step version: draft it yourself, adopt the one hard rule, turn off the flattery, run targeted editing passes, rewrite the flagged problems by hand, use the model only for the checks it is genuinely good at, compare versions with a fresh model instance, and learn the underlying craft so you can direct the tool instead of following it. If your goal is a first-person LinkedIn post rather than an essay, pair this with [writing authentic AI-assisted LinkedIn posts](/how-to/write-authentic-ai-assisted-linkedin-posts).

The steps

Write the whole first draft yourself. Start with your own words on the page, however rough. The model cannot improve a voice it never received, and the moment you ask it to generate the draft, the output is its prose with your topic bolted on, not your writing. Get the argument, the examples, and the ordering down first — bad sentences are fine, because fixing sentences is exactly the job you are about to hand off. What must be yours from the start is the thinking and the phrasing; the polish comes later and comes from you too.
Adopt the one hard rule: never use a word the model picks for you. This is Ptacek's Rule Number One, and it is the whole game: you may not use a single specific turn of phrase the LLM suggests, even when it sounds better than yours. "Better" is the trap — models are supernaturally good at pleasing phrasing, and adopting it is how a distinctive voice turns generic without you noticing. When a suggestion tempts you, treat it as a flag that the sentence needs work, then solve it in your own words. Use the model to locate problems, not to supply solutions.
Turn off the flattery — ask for problems, not praise. Left to its defaults, a model will tell you your draft is strong, which is both false and corrosive: praise makes you defensive about weak material you should be rewriting. Prompt against it explicitly — "do not compliment this; list only what is wrong" — and ask for specific, located criticism rather than a verdict. You want the passes to feel like a tough editor's margin notes, not a fan letter. If the output starts validating you, restate the instruction; sycophancy is the failure mode that quietly protects your worst paragraphs.
Run targeted editing passes, one flaw at a time. Instead of "make this better," run narrow passes the model can execute reliably: flag every passive construction, every filler word ("very," "really," "actually," "just"), every nominalization (a verb hiding as a noun — "make a decision" for "decide"), every repeated word, and any paragraph that would read better somewhere else. Each pass returns a list of locations, not a rewrite. Narrow prompts get accurate answers; broad ones get the model rewriting your voice, which is precisely what you are avoiding.
Rewrite every flagged problem in your own words. For each location the model surfaces, fix it yourself — turn the passive active, cut the filler, unpack the nominalization, break or move the paragraph. This is where the rule from step two earns its keep: you accept the diagnosis and reject the prescription. It is slower than clicking "accept," and that friction is the point — the result reads as a sharper version of you rather than a smoother version of the model. The tedious flaw-spotting was the model's; the writing stays yours.
Use the model only for the checks it's genuinely good at. Beyond editing passes, lean on the model for narrow, verifiable tasks: spelling and grammar, a sanity check on facts and names (which you then confirm against a primary source, because models invent citations confidently), and an occasional thesaurus when a word is on the tip of your tongue — noting that a word you retrieve with its help is different from a phrase it composes for you. Keep the boundary sharp: mechanical assistance and lookup are fine; generation of your sentences is not.
Compare versions with a fresh model instance. When you have a revised draft, paste the original and the rewrite into a new, context-free chat and ask which is stronger and why. A fresh instance has no memory of praising your earlier version, so it grades the text rather than the relationship, and the divergences it flags are worth a second look. Use its answer as a signal to reconsider specific passages — not as a scoreboard, and never as license to paste its wording back in. It is a second reader, not a co-author.
Learn the craft so you can direct the tool. The reason these passes work is that you know what a passive verb or a buried subject costs a sentence. Ptacek points writers to Joseph Williams's "Style: Lessons in Clarity and Grace" for exactly this vocabulary — clarity, cohesion, emphasis, concision — and it doubles as a source of better prompts — for instance, asking the model to flag sentences where the grammatical subject isn't the real actor. The more of the craft you own, the more the model becomes an instrument you play rather than a voice you borrow. That ownership is what separates writing with an LLM from being written by one.

Common gotchas

Accepting a suggestion because it 'sounds better.' That is the exact failure — pleasing model phrasing is what readers detect, so a better-sounding line the model wrote is worse than a rougher line you wrote.
Leaving the flattery on. A model that opens with 'great draft!' is training you to defend weak material; prompt for problems only, and restate it whenever praise creeps back in.
Asking for a rewrite instead of a diagnosis. 'Rewrite this paragraph' hands the model your voice; 'tell me what's wrong with this paragraph' keeps it. Always request located problems, not replacements.
Trusting the fact-check. Models state wrong facts and fabricate citations with total confidence — verify anything the model 'confirms' against a primary source before it ships.
Comparing drafts in the same long session. The model has already praised your earlier version and its context is biased; use a fresh instance for any old-vs-new comparison.
Over-editing until it's homogenized. The goal is a sharper you, not a smoother average — if the edits are sanding off the quirks that make the writing yours, you have edited too far.

Legal note

Using an LLM to copyedit your own writing is generally uncontroversial, but disclosure norms vary by context. Academic, journalistic, and some professional settings expect you to disclose AI assistance — and a few prohibit it — so follow the policy of wherever the writing is published. The workflow here (you write, the model flags flaws, you rewrite) keeps authorship clearly yours, which is the honest position most disclosure rules are asking you to be able to stand behind.

Where Kompozy fits

Be clear on where this guide and Kompozy part ways, because pretending otherwise would be dishonest. When the writing is a personal essay, an opinion piece, or the founder's own blog — anything whose value is a specific human voice — the advice above is the whole answer: you write it, the model copyedits, and no automation tool should own those sentences. Kompozy is not trying to be your ghostwriter for that work. Where it earns its place is the layer this guide barely touches: the recurring, multi-platform distribution content — the Text Posts, Blog Articles, Carousel captions, and Email Newsletters a brand needs every week across eight social platforms plus blog and email — where hand-drafting every asset is not craft, it is a bottleneck.

What's useful is that Kompozy operationalizes the two hardest disciplines from this workflow at that scale. Rule Number One — never let the model's pleasing phrasing through — is enforced by a [banned-word and banned-phrase filter](/glossary/persona-brief): the exact 'delve,' 'in today's landscape,' 'unlock,' rule-of-three tells that mark model prose are blocked at generation, so the drafts come out closer to your register instead of the headline voice readers detect. And the [Persona Brief](/glossary/persona-brief) governs voice across every asset, so ten posts from one source read as one company rather than ten different models — the consistency you would otherwise fight for by editing each piece by hand.

The copyeditor pass this article insists on still belongs to you, and Kompozy is built to keep it there. Every generated draft lands in a [per-post review pipeline](/glossary/autopilot) before anything publishes — that gate is exactly where you apply steps four and five to the distribution layer: reject the phrasing that slipped through, sharpen a weak hook in your own words, and confirm any fact the model asserted. You are not accepting machine prose wholesale; you are editing at volume with the tedious flaw-spotting already done. Use this guide's method for the writing that must be yours alone, and use Kompozy's brief-plus-review loop to hold the same standard across the content you cannot realistically hand-write one at a time. Starter ($99/mo, 5,500 credits) fits a solo writer running that review loop over a modest cadence; Pro ($299/mo, 18,000 credits) suits a creator or small team publishing across every platform each week; Enterprise is custom for agencies enforcing one voice across many brands.

Frequently asked questions

Can readers really tell when an LLM wrote something?

Often, yes — and more reliably than writers assume. Frontier models default to a uniform, polished register that is fine in isolation but recognizable in volume, and Ptacek's framing is that readers detect model words 'in the parts per trillion.' The point of writing your own draft and rejecting the model's phrasing is precisely to keep your text out of that detectable register. You cannot reliably prompt the tell away, so the safer move is to not generate the prose in the first place.

What's the single most important rule for writing with an LLM?

Never use a specific word or phrase the model suggests, even when it sounds better than yours. Ptacek calls this Rule Number One, and it is the difference between an LLM that sharpens your voice and one that replaces it. Let the model locate problems — passive voice, filler, weak structure — but solve every one of them yourself. A suggestion that tempts you is a signal the sentence needs work, not a phrase to paste in.

Is it cheating to write with an LLM, and do I have to disclose it?

Copyediting your own draft is not ghostwriting, and most people would not call it cheating — you did the writing; the model flagged flaws you fixed. Disclosure is a separate question that depends on context: academic, journalistic, and some professional settings expect (or occasionally forbid) disclosure of AI assistance, so follow the policy where you publish. The write-it-yourself workflow keeps authorship clearly yours, which is the honest answer to almost any disclosure rule.

Which model is best for editing my writing?

For copyediting, capability matters less than how you use it — any current frontier model (Claude, GPT, Gemini) can flag passive voice, filler, and structure competently. What matters is the workflow: narrow single-flaw passes rather than 'make this better,' flattery turned off, and a fresh instance for comparing drafts. Pick the model whose editing notes you find sharpest, and remember that a stronger model is a stronger editor, not a license to let it write.

Won't editing with an LLM make everything sound the same?

Only if you let it rewrite. Homogenization comes from accepting the model's phrasing, which trends toward one polished average voice. If you use it strictly to diagnose problems and you fix each one in your own words, the output stays distinctly yours — arguably more distinct, because you have cut the filler and passivity that blur any writer's voice. The over-editing risk is real, so stop when the edits start sanding off your quirks.