# How to Make ChatGPT Sound Less Like AI: Prompt + Edit Pass (2026)

> A prompt block to paste into ChatGPT, the settings that help and the one that doesn't, and a 15-minute edit pass that turns the draft into yours.

Source: https://hidengpt.com/guides/how-to-make-chatgpt-sound-less-like-ai

# How do I make ChatGPT sound less like AI?

Updated October 10, 2026

· 8 min read · by the HidenGPT team

Give ChatGPT what it can't guess (your facts, your reader and a sample of your writing), paste a prompt that rules out the habits it falls back on, then do one edit pass yourself. The prompt removes the obvious patterns. The edit pass is what puts you back in, because no setting can add a detail that nobody but you knows.

No prompt makes a draft read as human every time, and this page makes no claims about AI detectors. The goal is a draft you would sign without wincing.

| Line to paste | What it prevents |
|---|---|
| “Write this as me, in first person, for [who will read it]. After reading, they should [decision or action].” | The all-purpose voice that talks to everyone and asks nothing of anyone |
| “Use only these facts: [names, numbers, dates, examples]. Where a fact is missing, write [missing] instead of inventing one.” | Invented details, and filler like “many people” or “various factors” |
| “Match the sentence length and word choice of this sample of my writing: [150 to 300 words in the same genre].” | The default assistant register (it helps, though no model matches a style fully) |
| “Mix short and long sentences, including some under eight words.” | Flat rhythm |
| “Start with the point. No opening that announces the topic, no closing summary, and say each point once.” | Throat-clearing intros and recaps |
| “No lists of three unless I gave you three items, and no ‘not X, it's Y’ or ‘rather than’ contrasts.” | The two structural habits readers spot fastest |
| “Write ‘is’ and ‘has’, not ‘serves as’, ‘stands as’ or ‘boasts’. Don't end sentences with an ‘-ing’ clause that comments on the sentence.” | Inflated verbs and tacked-on commentary |
| “No em dashes. Use commas, colons, periods or parentheses.” | The best-known punctuation habit |
| “Avoid these words: delve, crucial, pivotal, tapestry, testament, landscape, foster, elevate, seamless, leverage, meaningful.” | The most recognizable stock vocabulary, old and new |
| “Plain words I would say out loud. Contractions are fine. No headings, bold or bullet points unless I ask.” | Report formatting in a message, an email or a post |
| “After the draft, under a line, list every sentence that relies on something I didn't tell you.” | Unchecked claims slipping into the final version |

## Before you paste anything

Gather three things first, because the prompt only works with them.

Your facts. Names, numbers, dates, the example you would use if you were explaining this out loud. A model with no facts fills the space with safe generalities, and those generalities are most of what makes a draft sound generated.

Your reader and what you want from them. “My landlord, who needs to book a plumber by Friday” gives the draft a job. “A professional audience” gives it nothing.

A sample of your own writing, about 150 to 300 words of the same kind of text (a rule of thumb, not a measurement). An old email of yours is a better sample for an email than your best essay. Keep expectations modest: in a 2025 PNAS study, models given 500 words of a human text and asked to continue in the same style still kept their own grammar habits, such as “-ing” clauses at 2 to 5 times the human rate. A sample moves word choice more than sentence shape, which is why the edit pass below exists.

## How to use the prompt

Put the lines from the table into one message, in that order, with your facts filled in. Then add the request on its own last line, for example: “Request: a 180-word email to my landlord asking for the boiler to be fixed by Friday.” Facts come before style rules on purpose: if the message runs long, the part the draft cannot do without is at the top.

For a quick reply you can cut it to three lines: the reader and goal, the facts, and “Plain words, varied sentence length, no em dashes, no summary at the end.”

Why a list of habits instead of “make it sound human”? A one-line request gives the model nothing specific to change, so it tends to trade one set of habits for another. Naming the habits gives it something to check against, and the last line of the prompt gives you a list of claims to verify.

## Settings that help, and one that doesn't

Custom Instructions. The style lines from the table, everything except your facts, sample and request, come to about 700 characters, which fits the 1,500-character limit OpenAI gives the Free and Go plans (paid plans get 5,000). On the web, they live under Settings, then Personalization. Saved there, they apply to every chat, so each new prompt only needs the facts, the sample and the request. OpenAI said in November 2025 that GPT-5.1 follows custom instructions and style preferences more reliably than earlier models.

Characteristics. Since December 2025 the Personalization pane has controls for warmth, enthusiasm, headers and lists, and emoji. Setting fewer headers and lists helps if you mostly write prose.

Base style and tone is the one that won't fix your drafts. OpenAI's help page says that when you ask for an email, a social post or a résumé, ChatGPT matches the tone to your instructions and the context, not necessarily to the personality you picked. Choosing Efficient or Candid changes how ChatGPT talks to you more than how your email reads.

## The edit pass (about 15 minutes for 600 words)

The time is a rule of thumb, not a measurement. Work in this order, because each step removes text the next step would otherwise waste time on.

1. Cut the opening. Read the first two sentences. If they set the scene or repeat your request back, delete them; the draft usually starts working at sentence two or three.

2. Clear the list under the line. Every sentence ChatGPT marked as resting on something you didn't tell it gets a real fact from you, or gets cut.

3. Swap general for specific. Find each “many”, “various”, “a range of” and “significant”, and put your number or example in its place, or delete the sentence.

4. Cut paragraph endings that repeat their paragraph.

5. Break the rhythm once per paragraph by splitting a long sentence or joining two short ones.

6. Read it aloud. Rewrite any sentence you wouldn't say to the reader's face, and add the one detail you would mention first in person.

7. Check names, numbers and dates against your notes. This comes last because the earlier edits move them around.

## When the second draft still sounds generated

Ask for one specific change at a time instead of a whole new draft. Requests that tend to work: “Delete the first paragraph and open with [fact].” “Rewrite paragraph 3 around my example about [example] and keep my wording where you can.” “Sentences 4 to 6 are the same length; make one of them under eight words.” “You used ‘rather than’ twice; state those points directly.”

Targeted requests matter more as models change. A September 2026 study by Graphite found that 55% to 72% of the patterns in one model's writing changed in its next version, so a fixed list of banned words goes stale while a habit you can name still works.

If you would rather start from a tool built for this, HidenGPT, an AI ghostwriter, drafts from your notes in a voice built from samples of your writing, then runs a red pen over the result and shows a naturalness score before and after. That score measures readability, not what a detector will say, and the last read is still yours.

## Common mistakes

- Asking it to “sound human” in one line: name the habits to drop and give it your facts.
- Pasting a word ban and nothing else: it reaches for near neighbors, so rule out the structural habits too.
- Using a sample from a different genre: a text message teaches it little about a cover letter, so give a sample of the same kind.
- Editing word by word from the top: fix the opening, the endings and the rhythm first, since those cuts remove many weak words anyway.
- Sending it with a [missing] placeholder or an unchecked number: search for brackets and check every figure before it goes out.

## Questions

### What prompt makes ChatGPT sound less like AI?

One that gives it your facts, your reader and a sample of your writing, and names the habits to drop: flat rhythm, lists of three, “not X, it's Y” contrasts, summary endings and em dashes. The table on this page is a version you can copy.

### How do I make ChatGPT sound more human?

Give it material only you have, then edit the result yourself. Specific facts and your own phrasing do more than any style setting, and a one-line “sound more human” request tends to swap one set of habits for another.

### Does changing ChatGPT's personality make it sound less like AI?

Not for drafts. OpenAI's help page says that when you ask for an email, a post or a résumé, ChatGPT matches tone to your instructions and the context, not necessarily to the personality you chose.

### Can I make ChatGPT write like me?

Partly. A 150 to 300 word sample of your writing in the same genre pulls its word choice toward yours, but a 2025 PNAS study found models keep their own sentence habits even when asked to continue a human text in its style, so plan to edit.

### Why does ChatGPT keep using the same phrases after I tell it not to?

Because instructions change what the model is likely to write rather than setting hard rules. OpenAI said in November 2025 that GPT-5.1 follows Custom Instructions more reliably, so put lasting rules there and keep checking drafts.

### How do I make ChatGPT emails sound less robotic?

Ask it to open with the reason you're writing, give it one concrete detail such as a date or an order number, and cap the length (under 150 words is a good rule of thumb for most emails). Add “no closing summary” and cut any pleasantry you wouldn't type yourself.

### Will this make my text pass an AI detector?

That isn't what this page is for, and nobody can promise a detector result. Detectors are unreliable in both directions, so aim for writing a reader trusts and follow your school's or employer's rules on AI use.

## Sources

- [OpenAI Help Center: ChatGPT Custom Instructions](https://help.openai.com/en/articles/8096356-chatgpt-custom-instructions) (checked October 10, 2026)
- [OpenAI Help Center: Customizing Your ChatGPT Personality](https://help.openai.com/en/articles/11899719-customizing-your-chatgpt-personality) (checked October 10, 2026)
- [OpenAI Help Center: ChatGPT release notes (12 November 2025, 19 December 2025)](https://help.openai.com/en/articles/6825453-chatgpt-release-notes) (checked October 10, 2026)
- [Reinhart et al., Do LLMs write like humans? Variation in grammatical and rhetorical styles, PNAS (2025)](https://pmc.ncbi.nlm.nih.gov/articles/PMC11874169/) (checked October 10, 2026)
- [Graphite research: AI Tells (16 September 2026)](https://graphite.io/five-percent/research/ai-tells) (checked October 10, 2026)

## Related

- [My draft reads like ChatGPT wrote it. How do I make it sound like a person?](https://hidengpt.com/write/make-ai-writing-sound-human)
- [I wrote this myself, but it sounds like AI. What am I doing that reads that way?](https://hidengpt.com/write/my-writing-sounds-like-ai)
- [I need a cover letter that sounds like me, not like ChatGPT wrote it.](https://hidengpt.com/write/cover-letter-that-doesnt-sound-like-ai)
- [I want LinkedIn posts that sound like me, not like a LinkedIn post generator.](https://hidengpt.com/write/linkedin-post-generator-in-your-voice)
- [How do I make ChatGPT write like me?](https://hidengpt.com/guides/how-to-make-chatgpt-write-like-you)
- [What custom instructions should I give ChatGPT for writing?](https://hidengpt.com/guides/chatgpt-custom-instructions-for-writing)
- [What words does ChatGPT use a lot?](https://hidengpt.com/guides/words-chatgpt-uses-a-lot)
- [What are the tells of AI writing?](https://hidengpt.com/guides/ai-writing-tells)
- [Why does ChatGPT use so many em dashes?](https://hidengpt.com/guides/why-does-chatgpt-use-em-dashes)
- [Does my writing sound like AI?](https://hidengpt.com/tools/ai-cliche-checker)

[Add HidenGPT as a preferred source on Google](https://www.google.com/preferences/source?q=hidengpt.com) to see these guides more often in search and AI Overviews.