How to make AI write like a human: 100 banned words and a ready prompt

Here is the set. It holds exactly 100 banned words and phrases in groups, plus limits and style rules. Copy the whole thing and paste it into your model's settings or the start of your task (where exactly, step by step, below). It is written as a direct command to the model.

Write in plain, human English. Do not use corporate filler or AI clichés. Follow the bans and limits below in every response.

Banned words and phrases, 100 items, never use them.

Inflated verbs, 15: delve, leverage, harness, utilize, streamline, foster, elevate, empower, embark, unlock, showcase, underscore, underpin, encompass, revolutionize.

Buzzword adjectives, 15: intricate, meticulous, pivotal, robust, comprehensive, seamless, multifaceted, commendable, invaluable, holistic, nuanced (as self-praise), cutting-edge, state-of-the-art, game-changing, unparalleled.

Buzzword nouns, 10: tapestry, landscape (figurative), realm, paradigm, synergy, ecosystem (figurative), framework (when it is not one), game-changer, deep dive, treasure trove.

Empty openers, 12: "In today's fast-paced world", "In the ever-evolving landscape", "In an era of", "Now more than ever", "It goes without saying", "As we all know", "It's no secret that", "Let's dive in", "Let's break this down", "Let's unpack this", "Imagine a world where", "Picture this".

Filler connectives, 12: "It's worth noting that", "It bears mentioning", "It is important to note", "Importantly", "Notably", "Interestingly", "Furthermore", "Moreover", "Additionally" (as filler), "That said", "As such", "In essence".

Empty closers, 9: "In conclusion", "To sum up", "In summary", "At the end of the day", "When all is said and done", "The bottom line is", "Ultimately", "All in all", "In a nutshell".

Sentence templates, 8: "It's not X, it's Y", "It's not just X, it's Y", "Not X. Not Y. Just Z", "The X? A Y.", "Whether you are X or Y", "From X to Y" with mismatched ends, "Despite these challenges", participle tails like "highlighting its importance / reflecting broader trends / contributing to".

Hedging, 5: "may help", "can be beneficial", "might prove useful", "in some cases", "to some extent".

Vague attribution, 9: "experts say", "studies show" without a citation, "research suggests", "it is widely believed", "many believe", "critics argue", "sources indicate", "according to reports", "data shows" without a source.

Promo inflation, 5: "world-class", "best-in-class", "unlock your full potential", "take it to the next level", "a testament to".

Limits per 1000 words: em dashes no more than two; three items in a row no more than one; lists no more than one; metaphors no more than two; rhetorical questions no more than one; one-sentence paragraphs no more than three. No emoji in articles. One summary, at the very end, not after every section.

Format: do not start list items with a bold label. Make headings declarative, not questions. Do not announce what you are about to do, just do it. Do not return to the opening question at the end.

Style: one name per object across the whole text, no synonym carousel. Vary sentence length: put an eight-word sentence next to a thirty-word one. At least one verifiable name, number, or date every 300 words.

That totals one hundred: 15 plus 15 plus 10 plus 12 plus 12 plus 9 plus 8 plus 5 plus 9 plus 5. This is the core. Below I break down each group with examples so you can adapt the list to your own work instead of pasting it blind.

Section 01Why AI writes like a robot in the first place

Understanding the cause helps you edit the set with intent. A model learns from enormous piles of text, and most of it is academic papers, corporate reports, manuals, and press releases. That is where "leverage," "furthermore," and "it's worth noting" live. The model absorbed that register as the norm of written English, because formal text outnumbers casual writing in its training by a wide margin.

The second cause is training on human ratings. The model was rewarded for answers that sound polite, careful, and inoffensive. That breeds endless hedging: "may be beneficial," "in some cases," "worth considering." Cautious mush draws fewer complaints than a plain claim, so the machine plays it safe. A human writes "this works," the machine writes "this may prove beneficial under certain conditions."

So a general request like "write naturally" does nothing: you are fighting what is baked into the model. Only a concrete set of bans and requirements overrides that default, not a polite plea to keep it simple.

Section 02The empty words AI uses instead of meaning

Start with the cleanest case: words you can delete right now, and the sentence gets better. Three kinds.

First, inflated verbs: delve, leverage, harness, utilize, streamline, foster, elevate, empower, embark, unlock, showcase, underscore, underpin, encompass, revolutionize. The machine writes "leverage our solution." A human writes "use our tool." "Utilize" is "use." "Foster growth" is "grow." Each of these verbs hides a plain action behind a layer of corporate padding, and the model needs to be told to drop them.

Second, buzzword adjectives with no measurement behind them: intricate, meticulous, pivotal, robust, comprehensive, seamless, cutting-edge, state-of-the-art, game-changing, unparalleled. The test is simple. "A robust tool" – robust how, in numbers? No answer means the word is empty. "A seamless experience" – seamless compared to what? Silence means you cut the adjective instead of hunting for a prettier one.

Third, buzzword nouns the model reaches for to sound profound: tapestry, landscape and realm in a figurative sense, paradigm, synergy, ecosystem, framework when it is not one, deep dive, treasure trove. "A rich tapestry of features" says nothing that "a set of features" does not.

Section 03Openers you can skip and still read fine

The model loves to warm up with an empty paragraph before it says anything. The openers: "In today's fast-paced world," "In the ever-evolving landscape," "In an era of," "Now more than ever," "It goes without saying," "As we all know," "It's no secret that." Add the false invitations "Let's dive in," "Let's break this down," "Let's unpack this," and the visual setups "Imagine a world where," "Picture this."

They share one thing: a signal that the author has not started talking yet. So the set tells the model to open on the point. Compare. The machine writes: "In today's fast-paced world, artificial intelligence is transforming how experts work." A human writes: "AI writes an expert a hundred headlines in a minute." The second is a fact you grab onto. The first is throat-clearing, and it gets cut in full, which only strengthens the opening line.

Section 04Filler connectives and empty closers

Next the set targets the glue the model uses to join paragraphs that have no logical link: "it's worth noting that," "it bears mentioning," "it is important to note," "importantly," "notably," "furthermore," "moreover," "additionally" as filler, "that said," "as such." The test is the same for all of them: remove the connective and the meaning holds. "It's worth noting that the price rose" and "The price rose" say the same thing. If two paragraphs will not connect without "furthermore," the problem is not the connective. One of the paragraphs is dead weight.

Closers are just as predictable: "in conclusion," "to sum up," "at the end of the day," "when all is said and done," "the bottom line is," "ultimately," "all in all," "in a nutshell." An empty closer restates what you already said and sends the reader back to the start. Good writing lands its last strong idea and stops. The period says it is over. That is why the set carries a separate rule: do not return to the opening question at the end.

Section 05The "it's not X, it's Y" template and other tells

Here the ban is on a structure, not a word. The model loves a handful of frames, and they give it away instantly.

The main one is the antithesis reframe: "it's not X, it's Y," "it's not just X, it's Y," "not X, not Y, just Z." "It's not a tool, it's a philosophy." "It's not just a course, it's a system." It sounds sharp and means nothing, because there is a pose where the content should be. The living replacement is a plain claim. Instead of "it's not a course, it's a system," write "after the course you sell differently." Meaning appears exactly where the pose used to sit.

Nearby sits the question with an instant answer: "The result? A failure." Once per piece, fine. Three times, it is a jackhammer. Two more frames: "From X to Y" with mismatched ends ("from a simple idea to a cultural shift"), and participle tails like "highlighting its importance," "reflecting broader trends," "contributing to growth." You can always cut those tails, and the sentence straightens out.

Section 06"Experts say": vague attribution

A separate red flag, and the set carries a hard line for it: a claim pinned to a nameless source. "Experts say," "studies show" with no citation, "research suggests," "it is widely believed," "many believe," "according to reports," "data shows." The rule allows no middle ground: either a name, an organization, a year, and a link, or the claim is deleted. Not softened, deleted. "Studies show video outperforms text" is nothing. Who studied it, when, on what sample? If the model cannot name the source, it does not know whether the claim is true or something it made up that sounds true. It builds convincing-sounding sentences beautifully, and that is the trap. The set also bans invented pseudo-terms – "the paradox of choice," "the engagement trap" – presented as established. A real term with a source is fine. A clever coinage minted on the spot is a tell.

Section 07One paragraph, before and after

Enough theory. Here is what the set does on a live example. This is what a model outputs with no rules, on the prompt "write about the value of AI for an expert."

Before: "In today's fast-paced world, AI is a pivotal tool for any expert. It's worth noting that it can leverage your workflow and unlock unparalleled efficiency. It's not just technology, it's a game-changer. Experts say it is the future. Ultimately, embracing AI opens up a world of possibilities."

Take it apart. "In today's fast-paced world" – an empty opener, cut. "Pivotal tool" – empty adjective plus empty noun. "It's worth noting that" – a filler connective. "Leverage" and "unlock unparalleled efficiency" – three banned words in a row. "It's not just technology, it's a game-changer" – antithesis plus buzzword. "Experts say" – attribution to no one. "Ultimately" – an empty closer. Out of forty words, zero of substance.

After, with the same model running on the set: "AI writes an expert a hundred headlines in a minute, builds a month of content in one sitting, and answers routine client questions while they sleep. I tested it myself: what used to take a day now takes half an hour." Thirty-four words. A concrete claim, a personal check, a number. Not one marker. That gap is the whole point of the set.

Section 08Quotas: em dashes, triads, lists, emoji

The second part of the set is devices that are not banned, only limited. Their power is in the dose. Count per 1000 words.

Em dashes, no more than two. A triad, three items in a row, no more than one; the second one in a piece already reads as a pattern. Lists, one per 700 words, and only for something truly listable: steps, parameters, a comparison. A list item with a bold label in prose, zero. Metaphors, no more than two, and each has to explain something, not decorate. A rhetorical question with no immediate answer, one. A one-sentence paragraph for drama, no more than three and never back to back. A question heading, one per article, the rest declarative. Emoji in articles, zero.

And a separate rule for summaries: one per piece, at the very end. Not after every section, no "here's what we'll cover" before a block and "here's what we covered" after it. The machine loves to recap itself three times. A human says it once and moves on. That is why the limits matter more than the bans: they keep the text out of both corporate mush and dry protocol.

Section 09What AI will not do unless you tell it

The third part of the set is content requirements. A regex cannot check them, the model will not do them unprompted, and they are exactly what separates living prose from dead.

One name per object across the whole text. The machine calls one thing a "dashboard," then an "interface," then a "panel," afraid to repeat, and gives itself away. Then a varied rhythm: an eight-word sentence next to a thirty-word one. The model's rhythm is flat, like ruled paper, so you set the variation by hand.

More to add for serious writing. A real opinion: a claim the reader could argue with, or the text is faceless. Honest caveats: a spot where you admit uncertainty is a strong human signal, and the machine states everything with equal confidence. No promo tone in a neutral piece, no inflated stakes where a post about prices ends with the fate of an industry. And the anchor: at least one verifiable name, number, or date every 300 words. Concrete detail is the enemy of generation, because you cannot invent it safely, and demanding it beats a dozen word bans.

Section 10How to paste the set into Claude, step by step

Now where to put it. I will use Claude as the example; the logic is the same in other models.

Open Claude, click your profile at the bottom left, go to Settings, and find the field that asks what preferences Claude should consider in responses. Paste the whole set from the box above. From that point the model writes cleaner in every new chat. If the field feels too short for the full set, paste the core – the word bans, openers, connectives, closers, and templates – and keep the full version with the quotas as a separate checklist. Or make a Project in Claude and drop the whole set into its instructions, so the rules apply inside that project without touching your other chats. In ChatGPT the same text goes into "Customize ChatGPT" or a project's instructions; in other models, into the start of the task.

The difference shows up at once. Before the set, on "introduce my course," Claude returns: "This course is a comprehensive solution that empowers you to master sales." After the set, on the same prompt: "In two weeks on this course you build your first funnel and run it on a real audience." The first you throw away, the second you keep reading.

Section 11How to check the set worked

Once you paste it, confirm it actually took. Give the model a task where filler used to creep in: "introduce my product in three paragraphs" or "write an intro for an article about launching a course." Then read the answer against five points. Opening: no warm-up like "in today's fast-paced world." Verbs: no "leverage," "utilize," "empower." Connectives between paragraphs: no "it's worth noting" or "furthermore." Ending: it does not restate itself with "in conclusion." Concreteness: real names, numbers, and facts appear instead of "many" and "often."

If a marker or two still slips through, that is normal, the model is not perfect. Say so plainly: "you broke the set here, rewrite it to the rules." Usually the second pass comes back clean. If markers pour in by the handful, the set did not load: check that it saved in settings and that you are in a new chat, not an old one from before the rules. And keep a quick living test on hand: read the answer aloud. If it sounds like a person talking, the set works. If it sounds like a manual, add the missing bans and try again.

Section 12How to adapt the set to your niche

The set here is a base, but tune it to your actual work. Every niche has its own clichés the model reaches for. Writing sales copy? Add "world-class service," "wide range of solutions," "our team of professionals," "personalized approach." Writing about tech? Cut "innovative solutions," "digital transformation," "next-generation." Running a personal blog? Ban the fake warmth: "dear friends," "from the bottom of my heart," "I truly hope." The mechanism is simple: for a week, collect the phrases that grate in the model's answers on your topic, and move them into the set. In a month you have a personal list tuned to your work, not a generic template. And add a couple of your own living phrases and a scrap of text in your voice, so the model does not just strip out the wrong thing, it picks up the right one.

Section 13One set for the whole team

The set is also easy to share. If you have writers, assistants, or editors working with AI, hand them all the same text to paste. Then every piece comes out in one style, instead of one writer scrubbing filler by hand while another ships raw output. Keep the set in a shared doc, and anyone on the team pastes it before they write. It doubles as a review checklist: a "furthermore" or three em dashes in a row in the finished draft means a rule was broken, back it goes. That ends arguments about taste. There is a list, there are limits, there is a yes or a no. One set saves hours of edits and holds the bar steady no matter who writes.

Section 14Update the set every quarter

The set ages. Developers deliberately scrub known tells, and new ones move in: "delve" spiked in frequency after ChatGPT launched, got flagged, and faded, while "gated" and "load-bearing" arrived. So refresh the set quarterly, and you can hand that to the model too. Turn on web search first – the toggle under the input box, free in Claude. Without it the model answers from memory and may invent sources. Then ask it to research online which signs mark AI-written text, in English specifically, using only real sources it finds, inventing nothing. Have it build a full list, grouped into categories – words, sentence structure, text structure, facts, formatting – with a sign, a reason, and an example under each, at least twenty-five items, no repeats, and a reread at the end to add what it missed. Move any phrase that shows up in three of five fresh generations into the set, and drop anything you have not seen in two cycles.

Section 15Honest note: there is no magic button

One sober point to close on, or the set becomes a con. There is no "zero percent AI" button. AI detectors are unreliable in both directions: a systematic review of fourteen tools found none above eighty percent accuracy, and on writing by non-native English speakers, false positives reach sixty-one percent. A detector routinely flags a living human essay as machine-made, and the reverse. So the goal is not to beat a detector. The goal is for the AI to write the way a human writes. A text that is clean against the set but carries one invented citation is worse than a slightly rough text with two "it's worth noting" and accurate facts. The set removes the machine flavor and teaches the model a normal style. Checking the facts is still on you, and there is no offloading that.