The Best Prompt to Make AI Writing Sound Human (Copy and Paste)

Copy and paste a proven humanize AI text prompt, plus the editing principles behind it so you can fix any AI draft fast.

Here is the short version: most "humanize this text" prompts fail because they tell the AI to sound human without telling it what actually makes writing sound like a machine. The prompt below fixes that. It gives the model a checklist of real AI tells to hunt and kill, not a vague instruction to "vary your tone."

Copy it, paste it into ChatGPT, Claude, Gemini, or whatever you use. Then read the rest of this guide so you understand why it works. A prompt you understand is one you can adapt when it gives you a mediocre result.

The prompt (copy and paste this)

You are a professional copy editor. Rewrite the following text so it reads like a person wrote it — not a language model.

Specific problems to fix:
- Remove filler transitions like "In conclusion," "Furthermore," "It's worth noting that," and "In today's fast-paced world."
- Cut em dashes. Use commas, periods, or a new sentence instead.
- Break up any list of three things that exists purely for rhythm (e.g., "fast, reliable, and affordable"). If you're going to list things, make sure each item earns its place.
- Replace passive voice where it hides the actor. Prefer "The team missed the deadline" over "The deadline was missed."
- Delete adjective stacks. If two adjectives could be cut without changing the meaning, cut them.
- Remove phrases like "it's important to note," "it goes without saying," "at the end of the day," and "moving forward."
- Cut any sentence that repeats what the sentence before it already said in different words.
- Replace abstract nouns with verbs. "Make a decision" → "decide." "Have an impact" → "affect."
- Fix zombie nouns (nominalizations). "The implementation of the strategy" → "implementing the strategy."
- If a sentence has a main clause and a participial tail (e.g., "..., highlighting the importance of X"), either make the tail its own sentence or cut it.

After rewriting, do NOT add an explanation or summary. Just return the revised text.

TEXT TO REWRITE:
[paste your text here]

That is the working version. The sections below explain each instruction, show a before/after example, and answer the questions I see most often about this kind of prompt.

Why most "sound more human" prompts don't work

The generic approach ("rewrite this to sound more natural") puts the burden on the model to guess what you mean. The model trained on millions of documents, many of them AI-generated. Its default idea of "natural" has already drifted toward patterns you're trying to escape.

The prompt above is specific about symptoms. It names "Furthermore" as a tell. It names em-dash overuse. It names participial tail clauses. When you name a thing, the model can target it. When you say "sound natural," the model improvises, and improvisation lands you back in the same place.

There is a second problem: generic prompts produce generic output. You end up with text that traded one set of AI habits for another. The rewrite feels slightly warmer but still has that frictionless, brochure quality. The humanizer prompt at /humanizer-prompt goes deeper on this: it addresses voice and specificity, not just surface tics.

Before and after: what the prompt actually does

Here is a real example. The original came from a ChatGPT draft for a project management blog.

Before:

"In today's dynamic business environment, it is crucial for teams to leverage agile methodologies in order to foster collaboration, enhance productivity, and deliver seamless results. It is worth noting that organizations that fail to adapt may find themselves at a disadvantage."

That paragraph has almost every major AI tell: "In today's dynamic business environment," "crucial," "leverage," "foster," "seamless," a tidy three-part list, and a hedging tail sentence that says nothing.

After (run through the prompt above):

"Agile teams tend to ship faster, and with less internal friction, than teams running on waterfall schedules. Companies that haven't made the switch are often the ones watching competitors release updates they're still planning."

The rewrite cuts all the buzzwords. It uses a specific contrast (agile vs. waterfall) instead of abstract praise. The second sentence has a point of view rather than a vague warning. It is shorter. It is also more useful to the reader.

The prompt didn't do this by magic. It did it by removing the scaffolding that AI drafts hide behind.

How to adapt the prompt for different contexts

The base prompt works for blog posts, web copy, and most business writing. For other contexts, you will want to modify it.

Academic writing

Add this instruction: "Match a first-person academic voice. Use hedged language where appropriate ('this suggests,' 'the data indicate') but remove hedging that is purely defensive ('it could be argued that,' 'it is possible that')."

Academic writing has its own register. The goal is not to remove all caution. Genuine epistemic humility belongs in research. The goal is to remove the performative caution that AI adds to cover its bets.

Email and Slack messages

Add: "Make this conversational and direct. Write at a reading level for a colleague, not a client proposal. One idea per sentence where possible."

Short-form writing is where AI hedging shows up most. An email that starts with "I hope this message finds you well" is a tell. A Slack message that ends with "Please don't hesitate to reach out if you have any further questions" sounds like a customer service bot. The prompt extension pushes against both.

Marketing copy

Add: "Replace benefit-stacking with one specific claim. If the original says the product is 'fast, powerful, and easy to use,' pick the one that matters most to this audience and demonstrate it with a number or an example."

Marketing AI drafts pile on adjectives because they're trained to be persuasive, and persuasion looks like enthusiasm. It doesn't. Persuasion looks like specificity. One concrete number beats three adjectives.

For a deeper look at how prompts differ across the major AI platforms, see humanizer prompts for ChatGPT, Claude, and Gemini compared.

The editing pass you still need to do

No prompt eliminates the need for a human read. What the prompt handles well: surface patterns, filler phrases, structural tics. What it cannot handle: knowing what your audience actually cares about, catching a claim that's factually off, or preserving a turn of phrase you deliberately used.

After running the prompt, do one pass with this checklist:

Read the first sentence. Would you say it out loud? If not, rewrite it by hand.
Find the longest sentence. Can you cut it in half? Usually yes.
Look for the word "very." Delete all of them.
Find any place you used "things" or "aspects" or "elements." Those are placeholder nouns. Name the actual thing.
Check the ending. AI rewrites often trail off into a summary or a question ("Do you want to learn more?"). Cut to the last sentence that says something real.

The checklist works because it gives you specific targets, the same way the prompt does. Vague edits ("make this better") produce vague results.

If you want to get into the mechanics of building these instructions from scratch, the guide on how to write a system prompt that strips out AI tells walks through it at the instruction level.

A note on AI detectors

People often ask whether running text through the humanize prompt will fool AI detectors. The honest answer: maybe, sometimes, depending on the detector and the original text.

Detectors look for statistical patterns: token probability, perplexity, burstiness. The prompt above changes the surface text, which does shift those scores. But detectors update, models update, and no prompt gives you a permanent guarantee.

More practically: if you're worried about a detector, you probably care about whether the writing sounds human to a person reading it. That is a better bar to aim for. A reader who can tell the text was AI-generated is a more consistent signal than any tool score.

FAQ

Can I use this prompt on any AI tool?

Yes. The instructions are plain English and work in ChatGPT, Claude, Gemini, and any other chat-based model. The results will vary slightly by model. Claude tends to preserve sentence structure more, while ChatGPT often makes bigger stylistic moves. But the core cleanup happens in all of them.

How long can the text be?

Most models handle up to a few thousand words per request without degrading. For longer pieces (a full white paper, a long-form article), split the text into sections and run each section separately. Running everything at once on a long document increases the chance the model loses the thread toward the end.

Do I need to run it more than once?

Sometimes. First pass gets most of the surface tics. A second pass, with a slightly different instruction (try adding "be even more direct and cut any sentence that doesn't add new information"), often catches what the first missed. Two passes on a 500-word draft takes under a minute.

Will this work on my own writing, not just AI drafts?

Yes, and this is an underrated use case. Human writers also accumulate habits: overusing the same transition, defaulting to passive voice under deadline pressure, hedging claims to avoid commitment. The prompt doesn't care where the habits came from. It fixes them either way.

Is there a version I can use as a system prompt?

The version at /humanizer-prompt is formatted as a system prompt you can paste into the "instructions" or "custom instructions" field of most AI tools. That way, every output you get already goes through the cleanup rules without a separate step.

The most useful thing to take from this page is not the prompt itself but the principle behind it: specificity beats vagueness in every instruction you give an AI. "Sound human" is a vague brief. "Cut em dashes, remove filler transitions, break nominalization" is a list of jobs to do. The model does jobs. Give it jobs.

Also look at a ChatGPT prompt that makes copy sound like you wrote it if you want a version built around voice-matching rather than just tic removal. Different goal, different approach.