Here is the prompt, so you can stop reading if that is all you came for.
Rewrite the text below so it reads as though a careful human wrote it.
Rules:
- Vary sentence length deliberately. Follow a long sentence with a short one.
- Cut every transition that announces itself: furthermore, moreover,
additionally, in conclusion, it is worth noting, delve, tapestry, landscape.
- Prefer concrete nouns to abstract ones. "The build broke" over
"challenges were encountered during the deployment process."
- Use contractions where a person would.
- Do not add enthusiasm the original did not have. No "exciting", no
"powerful", no "game-changing".
- Keep every fact, number, and claim exactly as written. Change the prose,
not the content.
- Do not add a summary paragraph at the end.
Return only the rewritten text.
TEXT:
"""
[paste here]
"""That one does most of the work. The rest of this explains why it works, what it cannot do, and which popular prompts are wasting your time.
Why "make this sound more human" does nothing
The most common prompt is some version of make this sound more human or rewrite this so it doesn't sound like AI. It reliably produces text that is slightly more casual and no less machine-like.
The reason is that the instruction has no referent. "Human" is not a style the model can target. Ask for it and you get the model's average of what human writing looks like, which is the same distribution the text already came from. You have asked it to move toward the centre while standing at the centre.
Prompts that work name a specific, checkable change: sentence length variance, a banned word list, a register, a point of view. The model can verify those against its own output. It cannot verify "human".
The five that work
1. The variance prompt
AI text has unusually even rhythm. Sentences cluster around the same length, and clauses are balanced where a person would be lopsided.
Rewrite the text with deliberately uneven rhythm. At least one sentence
under six words. At least one over thirty. Never three sentences in a row
of similar length. Do not change the meaning.This is the single highest-leverage instruction, because rhythm is the thing people notice before they can say why.
2. The banned-phrase prompt
Models reach for the same connective tissue. Naming it explicitly works better than asking for variety.
Rewrite the text. These words and phrases are banned: delve, tapestry,
landscape, realm, testament, moreover, furthermore, additionally, crucial,
vital, it is worth noting, in today's world, navigate the complexities,
unlock, elevate, seamless, robust, leverage (as a verb).
If a banned word is the only accurate one, rewrite the sentence so it is
not needed.The last line matters. Without it the model substitutes a synonym and keeps the structure, which changes nothing.
3. The voice-transplant prompt
The strongest technique on this list, and the least used. Give the model a sample of your own writing and ask it to match that, rather than to match "human" in the abstract.
Below are two samples. SAMPLE A is my own writing. SAMPLE B is text that
needs rewriting.
Study SAMPLE A for sentence length, vocabulary level, how often it uses
contractions, how it opens paragraphs, and what it does instead of
transitions. Rewrite SAMPLE B to match. Keep SAMPLE B's facts.
SAMPLE A:
"""
[400+ words of your own writing]
"""
SAMPLE B:
"""
[the AI text]
"""Four hundred words is roughly the floor. Below that the model infers too little and falls back on its defaults.
4. The register prompt (academic writing)
Academic prose has its own conventions, and "sound human" actively fights them. Asking for contractions and short punchy sentences in a literature review makes the text worse and more conspicuous, not less.
Rewrite in the register of a graduate-level academic paper in [field].
- Formal but not inflated. No throat-clearing openers.
- Hedge claims the way a careful researcher does: "suggests", "is
consistent with", not "proves" or "revolutionises".
- Vary sentence structure; avoid three consecutive sentences with the
same shape.
- Keep all citations and claims exactly as given. Add nothing.
- No concluding summary paragraph unless the original has one.If you are submitting work for assessment, check what your institution actually permits. Policies differ sharply, and "I used a tool to rephrase it" is not a defence at most of them.
5. The specificity prompt
AI writing is vague in a particular way: it describes categories where a person would give an instance.
Rewrite the text. Wherever it makes a general claim, replace it with a
specific one if the specific version is already implied. Wherever it
gestures at a category ("various factors", "a number of studies",
"many organisations"), either name them or cut the sentence.
Do not invent facts. If a vague statement cannot be made specific from
what is here, delete it rather than inventing detail.The final paragraph is not optional. Without it this prompt hallucinates confidently.
The three that don't
"Write like a human, not an AI." Covered above. No referent, no change.
"Add typos and grammatical errors." This circulates widely and is bad advice. It degrades the writing to fix a problem readers were not having, and detectors do not key on typographical perfection. You end up with text that is both error-ridden and still rhythmically machine-like.
"Write at a 6th-grade reading level." Produces short sentences uniformly, which makes the variance problem worse. Uniform simplicity reads as machine-generated just as reliably as uniform complexity.
What prompts cannot fix
Worth being straight about the limits, because most articles on this topic are not.
A prompt rewrites surface features. It does not add what the text is actually missing: a point of view, a specific example you lived through, a number you looked up, an opinion you would defend. Text that has nothing to say reads as hollow whether a person or a model assembled it, and no instruction set fixes that.
On detectors specifically: scores move, but they move unpredictably. Detector outputs are probabilistic, they disagree with each other on the same text, and they change without notice when models are updated. Anyone showing you a prompt with a guaranteed score is showing you one run on one detector on one day.
| What a prompt changes | What it doesn't |
|---|---|
| Sentence rhythm and length variance | Whether you have a point |
| Vocabulary and phrase choice | Factual specificity |
| Register and formality | Lived detail and genuine opinion |
| Structural predictability | Whether the argument holds |
The honest summary: prompts make AI text less conspicuous. They do not make it yours.
Chaining them
The prompts above are more effective applied in sequence than merged into one instruction. Models follow four or five constraints well and ten poorly.
A reasonable order:
- Specificity first, while the structure is still intact and vagueness is easy to spot.
- Banned phrases second, to clear the obvious tells.
- Variance last, because it reshapes sentences that the earlier passes already settled.
Run the voice-transplant prompt instead of all three when you have a good sample of your own writing. It subsumes most of what the others do.
When a prompt is the wrong tool
Prompting is the right approach when you are working in a chat window anyway, when the text is short, and when you want control over each decision.
It is the wrong approach at volume. Pasting a 2,000-word document through three sequential prompts, checking each pass for invented facts, is slower than it sounds — and the failure mode is quiet, because a model that drops a number while rewriting does not announce it. If you are processing text regularly rather than occasionally, that is the point at which a dedicated tool earns its place, mostly by being consistent about preserving the facts you asked it to keep.