2026-05-13·7 min read

How to Write With AI Without Sounding Like AI

The first time I let AI write something for me end-to-end, I posted it and immediately knew it was a mistake. Not because the post was bad — it was structurally fine. Three sections, clear argument, no typos. The problem was that anyone who'd read me before would know in the first paragraph that I hadn't written it.

It wasn't the facts. It was the rhythm. The pivots that started with "moreover" and "furthermore." The way every paragraph closed with a tidy summary. The slight tendency to use three adjectives where one would do. AI writing has a tell, and I'd just published a 900-word version of it.

I've spent the year since figuring out which prompt patterns kill the AI tell and which ones make it worse. I now use AI for almost everything I write — blog drafts, email replies, sales copy, script outlines. None of it sounds like AI. Not because I'm a great editor, but because the prompt does most of the work before the editor has to.

This is what actually moves the needle.

The AI tell isn't grammar — it's rhythm

Most "AI sounds like AI" advice online focuses on banned words. Don't say "delve." Don't say "tapestry." Don't say "navigate the landscape of." Useful, but cosmetic. You can find-and-replace your way past those and the writing still feels off.

The deeper tell is rhythm. AI writing tends to:

Pace every paragraph the same way. Setup, development, neat conclusion. Repeat. Repeat. Repeat. Human writing has uneven paragraphs — some are one sentence, some are seven, some end mid-thought.
Use parallel structures everywhere. "It's not just X, it's Y." "Not only A, but also B." Once is fine. Five times in a 600-word post is a giveaway.
Resist contradiction. AI by default writes the cleanest version of an argument. Humans hedge, double back, admit they were wrong about something earlier, change their mind in the middle of a sentence. Clean argument = AI tell.
Over-summarize. Every section ends with a one-sentence wrap-up of what the section just said. Humans usually trust the reader.

Most prompt fixes for "make it sound human" target the wrong layer. They tell the model to "write more casually" or "use a conversational tone." That gets you contractions and exclamation points, not actual rhythm changes.

Stop saying "make it sound human"

I used to end every writing prompt with some variation of "write in a natural, human, conversational tone." The output got marginally more casual. It also got more annoying — more "Look, here's the thing" openings, more rhetorical questions, more performative friendliness.

What changed everything was being specific about the negative space. Instead of telling the model what tone I wanted, I started telling it what patterns to avoid.

Here's the writing prompt I now reach for first:

I'll paste a draft. Rewrite it so it reads like a human wrote it, not AI. Specifically: - Vary paragraph length aggressively. Some one-sentence paragraphs. Some long ones. Never two paragraphs in a row that pace the same way. - Cut every sentence that summarizes what the previous paragraph just said. - Remove parallel constructions like "not just X, it's Y" unless it actually adds something. - If an argument has a real exception or a place where I'd hedge, write the hedge — don't smooth it over. - No "in today's world," no "navigate the landscape of," no "moreover," no "furthermore." Write it the way a thoughtful person actually writes when they don't have a thesaurus open. Direct, uneven, willing to be a little rough. Keep my arguments. Don't add new ones.

Category: writing. Tool: Claude. Status: favorite. The "don't add new ones" line at the end is critical — without it the model will helpfully invent additional reasoning that I never made.

This prompt isn't magic. It still gives me drafts I have to edit. But it gives me drafts where the editing is on the level of word choice, not surgically removing AI rhythm from every paragraph.

Feed it your old writing, not adjectives

The second pattern that changed my output: stop describing the voice you want and start showing it.

If you have any meaningful body of writing — old blog posts, email threads, Slack messages, anything — paste a chunk of it into the prompt. The model is dramatically better at imitation than at following abstract style instructions. "Write in a punchy, witty, irreverent voice" is a terrible instruction. "Here are 800 words I wrote last month, match the rhythm and choice of words" is a great one.

This is the prompt I use when I want to write something new from scratch:

I'm going to write a blog post about [TOPIC]. Below is one of my recent posts, attached as a voice sample. Match its rhythm, paragraph length variation, willingness to admit uncertainty, and specific way of constructing arguments. Don't copy the topic — just the voice. The new post should be ~1200 words. The argument is: [ARGUMENT IN ONE LINE]. The three sub-points are: [LIST]. Open with a concrete moment, not a thesis. End on a sharp line, not a summary. Voice sample: [PASTE 600–1000 WORDS OF YOUR OWN WRITING]

Category: writing. Tool: Claude. Status: favorite. Roughly half my published blog drafts now start this way. The output isn't publish-ready — but it's close enough that an hour of editing turns it into something I'm willing to put my name on.

The common mistake is pasting too little. A 200-word sample doesn't give the model enough to imitate. 600–1000 words is the sweet spot — long enough to show the rhythm, short enough not to bury the actual instruction.

Use the model as an editor, not a writer

The third shift was realizing that AI is a much better editor than it is a writer. The drafts I write myself, then edit with AI, are consistently better than the drafts AI writes for me, even when I'm tired and the writing is rough.

This is the editing prompt I use most:

I'll paste a draft. Don't rewrite it. Instead: 1. Find the single weakest sentence and quote it back to me with one specific suggestion. 2. Find the single weakest transition between paragraphs and tell me how to fix it. 3. Identify any paragraph that's doing two jobs and should be split. 4. Identify any two paragraphs that are saying the same thing and should be merged. 5. If the opening line could be sharper, propose one alternative. Don't comment on grammar. Don't suggest broad restructures. Don't be polite — if a section isn't earning its place, say so.

Category: writing. Tool: Claude. Status: favorite. The output from this prompt is more useful than any rewrite, because it preserves my actual voice and just fixes the spots where the voice is letting me down.

The "don't be polite" line matters. Without it, every AI editor has been trained to soften criticism into helpful-sounding suggestions. With it, I get answers like "the third paragraph repeats the second; cut it" — which is what I actually need.

What the AI tell really is

Here's the thing nobody talks about: AI doesn't sound robotic because it's bad at writing. It sounds robotic because it's been trained to write the *average* of all writing on the internet. The average is smooth, parallel, summary-heavy, hedge-free, and structurally tidy. That's the tell. Not specific words — averageness.

The prompt techniques that work all do roughly the same thing. They push the model away from the average and toward something specific — your voice, your rhythm, your willingness to be a little rough. The more specific the push, the less the output sounds like AI.

If you take one thing from this post: stop asking for "natural, human writing." Start describing what you want the writing *not* to do. Stop describing your voice in adjectives. Start showing the model 800 words of it. Stop letting the model write your drafts. Start letting it edit them.

The AI tell isn't a model problem. It's a prompt problem. Once you fix the prompt, the writing stops sounding like AI — even when AI wrote most of it.