Will these prompts make my writing undetectable?

No. Prompts change how a draft reads; detectors measure statistical properties of the finished text, and the two are only loosely related. Anyone selling a prompt as a way to beat detection is selling something they cannot deliver.

Why does every prompt list its limits?

Because a prompt library that only says what each prompt does is an advertisement. Knowing where one stops working is what stops you using it on the wrong task and blaming the output.

Which prompt should I start with?

If you have a draft already, the cutting prompts do the most per minute — compression removes the connective throat-clearing that makes drafts read as machine-written. If you are starting cold, use an outline prompt and write the prose yourself.

Can I use these for academic work?

Only where your institution permits AI assistance, and then disclose it. The outline-only prompt is the most defensible: the structure comes from a model and every sentence is yours, which is straightforward to describe in a disclosure.

Do these work on every model?

They are model-agnostic in shape. Prompts that ban specific phrasings work well everywhere. Prompts asking for analysis before writing work better on models that reason in steps. Expect to adjust wording rather than approach.

Can I reuse these prompts?

Yes — copy them, edit them, publish them. They were written for this page rather than adapted from a prompt pack, and no attribution is required.

Prompt library

Eighteen prompts written for one job: getting a first draft out of a language model that reads less like one. Grouped by essays, email, blog and LinkedIn.

Runs entirely in your browser — nothing is uploaded, and there is no account or limit.

Getting an argument instead of a summary

Write 600 words arguing a specific position on [topic]. State the position in the first sentence. Do not give balanced coverage — argue one side and name the strongest objection to it in the final paragraph without resolving it.

Why it works: Models default to surveying a topic because balance is safer. Naming a position first and forbidding balance forces commitment, which produces the uneven emphasis that summaries lack.
Where it stops working: You get one argument, which means you have to know which one you want. It is worse than useless if you have not read enough to have a position.

Stopping generic claims

Here are three sources with page references: [paste]. Write 500 words that make an argument using only these. Every claim must attach to one of them. If a point cannot be supported from these three, leave it out and say so at the end.

Why it works: Constraining the evidence removes the model's ability to reach for plausible-sounding general knowledge, which is where most vague, machine-typical sentences come from.
Where it stops working: The model can still misattribute a claim to the wrong source. Check every citation against the original — this reduces invention, it does not eliminate it.

Matching how you actually write

Here are 300 words I wrote: [paste]. Describe my sentence rhythm, average sentence length, and three habits I repeat. Then write 400 words on [topic] using those habits. Do not smooth them out.

Why it works: Asking for the analysis before the writing makes the model commit to explicit features rather than a vague impression of your style, and gives you something to check the output against.
Where it stops working: It copies surface habits, not judgement. The structure of the argument will still be the model's, and that is usually the more obvious tell.

Using AI without it writing your prose

Do not write any prose. Produce an outline for [topic] as a numbered list of claims, one line each, in the order they should appear. For each, note in brackets what evidence it would need.

Why it works: The most defensible use in academic work: the structure comes from a machine and every sentence is yours. It also produces the strongest disclosure — you can state exactly what the tool did.
Where it stops working: Nothing about the outline is original. If the ordering is obvious, that is because the model reached for the obvious ordering.

Removing the padding a model adds

Cut this to 70% of its length without losing any claim: [paste]. Remove transitions that only announce what the next sentence does. Do not add anything. Show only the shortened version.

Why it works: Compression targets exactly the connective throat-clearing that makes drafts read as machine-written, and it is far more effective than asking a model to 'sound human'.
Where it stops working: Aggressive cutting can drop qualifiers that were doing real work. Read for changed meaning, not only for length.

Email that sounds like a person who knows the recipient

Write an email to [name], who I [relationship and last interaction]. I need [outcome]. Reference the last thing we discussed. Under 120 words. No opening pleasantry beyond a greeting, and no closing line offering further help.

Why it works: The two most machine-typical parts of an email are the warm-up and the sign-off. Banning both, and anchoring to a shared specific, removes most of the tell.
Where it stops working: It cannot know your history — anything you do not supply gets invented or omitted.

Saying no without sounding like a policy document

Write a short email declining [request]. Give the real reason in one sentence: [reason]. Do not apologise more than once, do not offer alternatives I have not listed, and do not soften the decision into ambiguity.

Why it works: Models hedge bad news until it reads as undecided. Capping the apology and forbidding invented alternatives keeps the decision legible.
Where it stops working: Directness reads differently across cultures and seniority. Adjust before sending.

Chasing without the stock phrasing

Write a follow-up about [thing], sent [n] days after the last message. Do not use the words 'just', 'circling back', 'touching base', 'checking in' or 'wanted to'. Give one new piece of information that makes replying easier.

Why it works: Follow-ups are the most formulaic email genre, so banning the formula by name works better than asking for originality.
Where it stops working: If you have no new information, the prompt will manufacture the appearance of some.

Cutting a draft you already wrote

Rewrite this email at half the length, keeping my phrasing wherever possible: [paste]. Change only what has to go. List anything you cut that carried meaning.

Why it works: 'Keep my phrasing' preserves the specific word choices that mark writing as yours, while the audit line stops silent removal of things that mattered.
Where it stops working: Half is arbitrary. Some emails are already as short as they can be.

An opening that is not throat-clearing

Write the first 120 words of an article about [topic]. Open with a specific case, number or moment — not a definition, not a statement about the modern world, not a rhetorical question. The reader should learn something factual in the first sentence.

Why it works: Model openings default to scene-setting abstraction. Naming the three failure modes and requiring a fact up front removes them more reliably than asking for a 'strong hook'.
Where it stops working: It will invent a specific if you do not supply one. Give it the case or check what it produces.

Breaking the even-sentence pattern

Rewrite this section so sentence length varies sharply: [paste]. At least two sentences under six words. At least one over thirty. Never three consecutive sentences of similar length. Do not change the meaning.

Why it works: Uniform sentence length is one of the most visible machine tells. Specifying the distribution numerically works where asking for 'natural rhythm' does not.
Where it stops working: Overdone, this reads as affected. Use it on a section, read it aloud, and revert anything that sounds staged.

Replacing abstractions with specifics

Go through this draft and mark every abstract noun standing in for something concrete — landscape, ecosystem, journey, framework: [paste]. For each, ask me what specific thing I meant. Do not guess.

Why it works: 'Do not guess' turns the model into an editor asking questions rather than a writer filling gaps with plausible detail, which is where invented specifics come from.
Where it stops working: You have to answer the questions. It surfaces the work rather than doing it.

Adding the objection you are avoiding

Read this draft: [paste]. Write the strongest objection a knowledgeable sceptic would raise, in 100 words. Do not soften it and do not answer it.

Why it works: Drafts read as machine-made partly because they are frictionless. An unanswered objection adds the texture that agreement-shaped writing lacks.
Where it stops working: The objection is only as good as the model's grasp of the field. In specialist areas it will produce the obvious one.

Finding where a long piece sags

Summarise each paragraph of this draft in five words: [paste]. Then list any paragraphs whose summaries are near-duplicates, and any place the sequence repeats itself.

Why it works: Repetition across 2,000 words is hard to see while reading and obvious in a list of summaries. This is diagnosis rather than rewriting, so nothing gets silently changed.
Where it stops working: It finds structural repetition, not tonal flatness. Those are different problems.

A post that does not read like every other post

Write a 150-word LinkedIn post about [topic]. No single-sentence paragraphs stacked for effect. No opening one-line hook. No question at the end. No emoji. Write it as continuous prose, as though for a colleague.

Why it works: LinkedIn's house style is itself the machine tell, because models learned it from the platform. Banning the format by name is the fastest route out of it.
Where it stops working: It reads unlike the feed, which is the point and also a reach trade-off you are choosing.

A story post without the fake epiphany

Here is what happened: [paste facts]. Write 180 words about it. Do not add a lesson, do not generalise to 'what this taught me about leadership', and end on the last concrete detail rather than a reflection.

Why it works: The manufactured takeaway is the most recognisable pattern in AI-written social posts. Forbidding it leaves the detail, which is the part worth reading.
Where it stops working: Some posts genuinely need a point. This prompt refuses to supply one.

Replies that add something

Here is a post: [paste]. Write a 40-word reply that adds one specific fact, disagreement or example the post does not contain. If I have nothing to add, say so instead of writing a reply.

Why it works: The escape hatch matters more than the prompt. Most AI comments are empty because the model was required to produce something; permitting silence removes that pressure.
Where it stops working: It will still sometimes produce a reply when 'nothing to add' was the honest answer.

Tightening your own post without losing your voice

Tighten this post without changing my phrasing: [paste]. Only remove words and merge sentences. Do not introduce vocabulary I have not used. Show a diff of what you cut.

Why it works: Restricting the model to deletion makes it impossible to overwrite your voice, which is what happens when you ask for a general rewrite.
Where it stops working: It cannot fix a post whose problem is that it has nothing to say.

What prompts can and cannot change

A good prompt changes what the model writes about, how long the sentences are, and which stock formulations it avoids. Those are real improvements and they are visible to any reader.

What a prompt does not meaningfully change is the statistical signature detectors measure. Perplexity and burstiness are properties of the finished text as a whole, and constraining vocabulary or banning phrases moves them far less than the marketing around "undetectable prompts" suggests. The mechanism is explained here.

We sell a rewriter, so treat this as an interested party being straight with you: HumanFlow does not promise to beat any detector, and neither does any prompt on this page. What both can do is make a draft read more like a person wrote it, which is worth doing on its own terms.

The four moves that do most of the work

Ban the formula by name. Asking for "natural writing" produces nothing. Listing the exact phrases to avoid — "circling back", single-sentence paragraphs, the closing question — works immediately, because the instruction is checkable.

Constrain the evidence. Giving the model three sources and forbidding anything outside them removes its ability to reach for plausible general knowledge, which is where most vague sentences originate.

Ask for deletion, not rewriting. "Cut this to 70% without losing a claim" preserves your phrasing while removing padding. A general rewrite replaces your voice with the model's.

Give it permission to refuse. Several prompts here allow the model to say it has nothing to add. Most empty AI output exists because the model was required to produce something.

How to tell whether a prompt worked

Judge the output, not the prompt. Four checks take about a minute between them and catch most of what goes wrong.

Read it aloud. Machine-typical prose is hard to read aloud naturally because the sentences are all the same shape, so you run out of breath in the same place repeatedly. This is the fastest test and the one most people skip.

Look for the manufactured takeaway. If the last paragraph generalises from a specific into a lesson nobody asked for, the model added it. Cut it and check whether the piece is worse — it usually is not.

Count how many sentences you would defend. Not agree with; defend, to someone who pushed back. Frictionless drafts tend to score badly here, and the fix is usually to add the objection rather than to rewrite the prose.

Check whether anything is invented. Prompts that constrain evidence reduce fabrication; none eliminates it. Every statistic, quote and citation needs verifying against the source regardless of how the draft was produced.

Writing your own

Every prompt here is built from the same four moves, and once you see them you can stop collecting prompts and write the one your task needs.

Name what to avoid, specifically. Abstract instructions produce nothing — "write naturally", "sound human", "be engaging" are all unfalsifiable, so the model cannot tell whether it complied. "Do not use the words just, circling back, or touching base" is checkable, and checkable instructions get followed.

Constrain the inputs. Supplying the sources, the facts or the outline removes the model's room to reach for plausible general knowledge, which is where both vagueness and invention come from.

Restrict the operation. "Only delete", "only merge sentences", "do not introduce vocabulary I have not used" keep your voice intact. An unrestricted rewrite returns the model's voice every time.

Permit refusal. Adding "if there is nothing to add, say so" removes the pressure to produce, which is the source of most empty output. It is the single most underused line in prompting.

Using these honestly

Where AI use is permitted and disclosed, these prompts help you get a better draft faster. Where it is banned, no prompt makes it acceptable — and the outline-only prompt exists precisely because it produces a use you can describe in one sentence to an instructor.

Once you have a draft, the AI word finder shows which phrases survived, and the sentence rhythm tool shows whether the lengths still vary.

Common questions

Will these prompts make my writing undetectable?: No. Prompts change how a draft reads; detectors measure statistical properties of the finished text, and the two are only loosely related. Anyone selling a prompt as a way to beat detection is selling something they cannot deliver.
Why does every prompt list its limits?: Because a prompt library that only says what each prompt does is an advertisement. Knowing where one stops working is what stops you using it on the wrong task and blaming the output.
Which prompt should I start with?: If you have a draft already, the cutting prompts do the most per minute — compression removes the connective throat-clearing that makes drafts read as machine-written. If you are starting cold, use an outline prompt and write the prose yourself.
Can I use these for academic work?: Only where your institution permits AI assistance, and then disclose it. The outline-only prompt is the most defensible: the structure comes from a model and every sentence is yours, which is straightforward to describe in a disclosure.
Do these work on every model?: They are model-agnostic in shape. Prompts that ban specific phrasings work well everywhere. Prompts asking for analysis before writing work better on models that reason in steps. Expect to adjust wording rather than approach.
Can I reuse these prompts?: Yes — copy them, edit them, publish them. They were written for this page rather than adapted from a prompt pack, and no attribution is required.

Need more than a utility?

HumanFlow rewrites AI drafts to read naturally and shows a sentence-level detection score — free up to 1,500 words per scan.

Try the humanizer free

Other free tools

Word & character counter Case converter Text cleaner Sentence rhythm analyzer Text diff checker Word frequency counter Character counter with platform limits AI word finder

← All free tools