Rewrite an AI draft so it stops reading like one, by putting back what the model left out
A rewrite prompt that finds the sentences giving your draft away, and fixes them by adding specifics rather than swapping words for synonyms.
- Works in
- ChatGPT, Claude
- You need
- An AI written draft you want to fix · One real detail only you know · Who the reader is
- Written for
- ai humanizer prompt
Scored by our own engine
This page, run through the audit we sell. Measured 8 August 2026.
Most tools sold as AI humanizers work at the level of vocabulary. They swap words, vary sentence lengths, and hand back text with the same shape and the same emptiness. The prompt below fixes the cause instead: it finds the passages that read as generated, names what specific thing is missing from each, and rewrites them around details you supply.
Why generated text reads as generated
It is not word choice. It is that the writer had nothing in particular to say and no particular person to say it to.
Ask a model to write about onboarding and it produces sentences that are true of every company’s onboarding. That is not a style failure, it is the correct output for the question. There is no number in it because no number was supplied, no example because no example was supplied, and no position because taking one would be a guess. The prose is a faithful reflection of an empty brief.
This is why synonym swapping never fixes it. You can replace every instance of “leverage” and the paragraph still says nothing that could be disagreed with.
The diagnosis is worth more than the rewrite
Running this prompt for the first time on a draft you were reasonably happy with is uncomfortable, which is the point. The ten quoted sentences in pass 1 are usually recognisable as habits rather than accidents.
The categories that catch the most, in our experience of running it: empty scaffolding, meaning a sentence whose only job is to introduce the next one, and no stake, meaning a passage that nobody could disagree with. A piece made mostly of those two is readable, professional and entirely forgettable, and it is what most content briefs actually produce.
The gap markers, and why they are ugly
Pass 2 writes [NEED: the figure for X] into the draft rather than filling it in.
That is deliberate and it is the constraint most people delete. A model asked to complete a sentence like “teams that do this see a [NEED: percentage] improvement” will happily supply a percentage, and it will be plausible, and it will be invented. The marker is left visibly broken so it cannot survive a careless copy and paste into a CMS.
If you find yourself with fifteen markers, that is not a failed run. It is an accurate measurement of how much of the draft was made of assertions you cannot support.
On detectors, plainly
A large share of the people searching for this want text that passes an AI detector, so it is worth being direct rather than quietly selling around it.
Detectors are unreliable in both directions. They flag writing by people who learned English as a second language and anybody writing in a formal register, and they miss machine written text routinely. There is no prompt that makes output reliably undetectable, and any product claiming otherwise is selling a coin flip with a confident interface. The measurements behind that, and what to aim at instead of a detector score, are in how to humanize AI content.
What you can influence is whether the piece is worth reading. That is a lower ceiling than “undetectable” and it is the only one that holds up.
Where this fits with writing for assistants
There is a convenient overlap. The qualities that stop a draft reading as generated, which are concrete claims, real examples and a passage that answers one question directly, are the same qualities that make a passage liftable into an assistant answer. Writing that hedges everything is both dull to read and useless to quote.
The pairing to run afterwards is the article draft prompt, which applies the same principle earlier, at the drafting stage, by refusing to write until you supply the specifics. Fixing it before it is written is cheaper than fixing it after.
You are an editor. I will paste a draft that a language model wrote. Your
job is to find what makes it read as generated and fix it, in that order.
Work in three passes and show me each one.
Pass 1. Diagnose before you touch anything.
Quote up to ten specific sentences or phrases from my draft that give it
away, and for each one name the reason from this list:
- Hedging: says something is important without saying what to do.
- Empty scaffolding: a sentence whose only job is to announce the next
sentence.
- Triad padding: three adjectives or three examples where one carried the
meaning.
- Unearned authority: a claim of significance with no number, source or
example behind it.
- Symmetry: every paragraph the same length, every section the same shape.
- Vocabulary tell: a word this reader would not use, listed in my style
notes below or obvious from the context.
- No stake: a passage that could not be disagreed with, because it does not
actually claim anything.
Pass 2. Ask what is missing, then rewrite.
For each diagnosed passage, say what concrete thing would fix it: a number,
a named example, a failure case, an objection, a date, a decision somebody
made. Then rewrite the passage using what I gave you in the details below.
Where I have given you nothing that fits, write the sentence with a clearly
marked gap like [NEED: the figure for X] rather than inventing a
plausible detail. An invented specific is worse than an obvious hole.
Pass 3. Deliver the rewritten draft in full.
Rules that decide whether this works:
- Do not swap words for synonyms to change the texture. Replacing
"leverage" with "utilise" changes nothing about why the draft reads as
generated. The problem is missing content, not vocabulary.
- Do not add errors, casual filler, or fake hesitation to sound human.
Deliberate sloppiness reads as a different kind of fake.
- Do not lengthen. If the rewrite is longer than the original, you have
padded rather than fixed. Aim for shorter.
- Vary sentence length because the meaning varies, not on a schedule.
- Keep every factual claim my draft made. If you think one is wrong or
unsupported, list it at the end under Claims I would check. Do not
silently soften it.
- Do not tell me the result will pass an AI detector. You cannot know that
and neither can I.
- Do not ask me questions before you start. State any assumption you make,
label it, and produce all three passes. Put your questions at the end.
My reader: [WHO THE READER IS]
Real details I can vouch for: [SPECIFICS ONLY YOU KNOW]
Words and phrases I never use: [STYLE NOTES, OR WRITE "none"]
My draft:
[PASTE THE DRAFT]What to change
Everything in square brackets is yours to replace. Nothing else needs editing.
[WHO THE READER IS]- One sentence on who reads this and what they already know. The single biggest cause of generated sounding prose is a draft written for nobody in particular, because a model with no reader in mind defaults to explaining everything at the same level to everyone.
[SPECIFICS ONLY YOU KNOW]- The part that does the actual work. A number you measured, a customer who said something, a thing you tried that failed, a date, a decision and why you made it. Three of these will change a draft more than any rewrite instruction. Leave it empty and the prompt will hand you back a cleaner version of the same hollow piece, with the holes marked.
[STYLE NOTES, OR WRITE "none"]- Words you never want to see. Most people have a list once they think about it: delve, tapestry, landscape, seamless, robust, in today fast paced world. Adding them here catches the vocabulary tells in one pass instead of over three rounds of editing.
[PASTE THE DRAFT]- Paste the whole thing rather than a section. The symmetry and scaffolding problems are only visible across a full piece, and a model shown two paragraphs will diagnose sentence level tics and miss the structural reason it reads flat.
How to run it
- 01Collect three real details before you open the model
Write down a number, an example and an objection from your own knowledge of the subject. This takes five minutes and it is the whole exercise. Everything the prompt does afterwards is arranging what you supply here, and a run with an empty details field produces a tidier draft that is just as anonymous.
- 02Run it and read pass 1 on its own
Stop after the diagnosis and read the ten quoted sentences before looking at any rewrite. That list is a description of your own writing habits, or of the model you drafted with, and it is more useful the second time than the first because you start avoiding those shapes at the drafting stage.
- 03Fill in every marked gap yourself
The rewritten draft comes back with [NEED: ...] markers where the prompt refused to invent something. Fill those in or delete the sentences around them. Publishing with the markers still in is the obvious failure, and publishing after asking the model to fill them in is the quiet one.
- 04Read it aloud once
Anything you stumble over is a sentence that survived the rewrite without earning it. This catches what no prompt catches, takes two minutes, and is the only step in this list that a tool cannot do for you.
- 05Check it can still be quoted
Run the published page through the AI content readiness check. Rewriting for a human reader occasionally breaks the thing that made a passage liftable into an assistant answer, usually by turning a direct answer into a lead in. Worth confirming rather than assuming.
Questions people ask
What is an AI humanizer prompt?
An instruction set that rewrites machine written text so it reads as though a person with something to say wrote it. The useful ones work by identifying what the draft is missing, which is usually specifics, and adding them. The weak ones swap words for synonyms and change the texture without changing the reason the draft felt empty.
How do I make ChatGPT write like a human?
Give it something only you know. Models produce generic prose because they are asked generic questions, and the fastest fix is to supply a real number, a named example and a real objection before asking for a single sentence. Style instructions help at the margin. Missing content is the actual cause and no rewrite instruction fixes it.
Will this make my text pass an AI detector?
We do not claim that and you should distrust anybody who does. AI detectors are unreliable in both directions: they flag human writing, particularly from people writing in a second language or in a formal register, and they miss machine writing routinely. This prompt is aimed at whether a reader finds the piece worth their time, which is the thing you can actually influence.
Why does the prompt refuse to invent details?
Because an invented specific is the worst possible outcome of a rewrite. A vague draft is merely dull, while a draft containing a plausible fabricated statistic is a credibility problem you will not notice until somebody else does. Marked gaps are ugly on purpose so they cannot be published by accident.
Does humanising AI content help with SEO?
Indirectly, and not for the reason usually given. Google states it rewards helpful content regardless of how it was produced, so the production method is not the ranking factor. What matters is that specifics, a real position and genuine experience are also the things that make a passage worth quoting, and both search results and assistant answers select for the same qualities.
Can I run this on a draft I wrote myself?
Yes, and the diagnosis pass is often more useful on human writing. Hedging, empty scaffolding and passages with no stake are not machine specific habits. They are what anybody writes when they are unsure of the audience or writing to a length target.
Unsubscribe in one click. We never pass your address on.