Prompts to Humanize SI (AI) Text: 5 Copy-Paste Prompts That Actually Work

Five tested prompts to make SI (AI) drafts sound human in ChatGPT, Claude or Gemini, the follow-up prompts for round two, and where prompting stops helping.

A smiling woman typing on a laptop at a sunny desk with sticky notes, a speech bubble with lines of text floating above the screen
SI-generated illustration

The best prompt to humanize AI text is not one magic sentence. It is a short brief that tells the model who is reading, which words to drop, how the rhythm should feel, and that it may not add a single new fact. Below are the prompts we actually use with ChatGPT, Claude and Gemini, ready to copy, plus the editing prompts for round two, and an honest look at where prompting stops helping.

Why "make this sound human" rarely works

Ask any chatbot to "make this sound more human" and you usually get the same text in a slightly different costume. The model swaps "utilize" for "use," adds a friendly opener, and keeps the rest: the three-part lists, the tidy closing line, the sentences that all run about the same length. It does not know what "human" means for your reader, so it guesses, and its guess is its own house style.

The fix is specificity. A good humanizing prompt names the audience, the register, the patterns to cut and the patterns to keep. Anthropic's own prompting guide makes the same point in general terms: it says that providing "context or motivation behind your instructions" helps the model deliver more targeted responses, and that examples are "one of the most reliable ways to steer Claude's output format, tone, and structure." That advice carries over to ChatGPT and Gemini too.

If you are not sure which patterns give a draft away, read our list of signs of SI writing first. You will write better prompts once you can spot the problem yourself.

Prompt 1: set the voice and the reader

Start every humanizing pass by telling the model who the text is for and who is supposed to be talking. This one prompt does more than any word list.

Rewrite the text below for [audience, e.g. small business owners who skim on their phone]. The writer is [who you are, e.g. an accountant with ten years of client work]. Write the way that person would explain it to a client across a desk: plain words, direct statements, no hype. Keep the same meaning and the same order of ideas. Text: [paste]

Fill in the brackets honestly. "An accountant with ten years of client work" pulls the vocabulary toward concrete terms. "A friendly expert" pulls it toward nothing in particular. The more real the persona, the less generic the output.

Prompt 2: a banned-word list, paired with what to do instead

Lists of forbidden words are popular, and they help, but only halfway. Anthropic's guide advises telling the model what to do instead of what not to do, and in our experience that is exactly where bare ban lists fail: the model drops "delve" and reaches for "dig into," which is the same tic with new paint. So pair the list with a positive rule.

Edit the text below. Do not use these words or phrases: delve, moreover, furthermore, crucial, pivotal, seamless, robust, leverage, landscape, tapestry, testament, "in today's fast-paced world", "it is important to note", "in conclusion". When you would reach for one of them, use the plainest word that says the same thing, or cut the phrase if the sentence works without it. Do not replace them with near synonyms. Text: [paste]

Our full list with plain replacements lives in SI words to avoid. Keep your own list short, though. Twenty banned words is enough; two hundred makes the model stiff and it starts avoiding ordinary words too.

Prompt 3: fix the rhythm

Rhythm is the tell most people feel but cannot name. Machine drafts tend to use sentences of similar length and similar shape, one after another. This prompt targets that directly.

Vary the sentence length in the text below. Mix a few very short sentences (under eight words) with longer ones. Do not start two sentences in a row the same way. Remove "Firstly", "Secondly" and "Additionally" scaffolding and let the ideas connect on their own. Do not add lists or headings that are not already there. Keep every fact. Text: [paste]

Watch the punctuation in the result. Models love the em dash () and will often add more of them while "varying rhythm." If that bothers you, add one line: "Use commas, periods or parentheses instead of em dashes." We explain why the habit exists in why SI uses em dashes.

Prompt 4: keep the facts, add nothing

This is the prompt that protects you. When a model is asked to make text more lively, it often invents color: a statistic, an example, a quote that sounds plausible. In a blog post that is embarrassing. In a report, a grant application or a client email it can be a real problem.

Rewrite for tone only. Keep every number, name, date, citation and claim exactly as written. Do not add examples, statistics, quotes or details that are not in the original. If a sentence is vague, leave it vague and mark it with [needs detail] instead of inventing something. Text: [paste]

The [needs detail] marker is the useful part. It turns the model into an editor that flags weak spots for you to fill with real information, instead of a writer that fills them with fiction.

Prompt 5: teach it your voice with a sample

A sample of your own writing beats any description of it. Paste 200 to 400 words you wrote without help (an old email, a post, a report section) and ask the model to match it.

Here is a sample of my writing: [paste your sample]. Notice my sentence length, how formal I am, whether I use contractions, and how I open and close paragraphs. Now rewrite the draft below so it reads like the same person wrote it. Do not copy phrases from the sample; match the style, not the content. Draft: [paste]

Two practical notes. Pick a sample from the same genre as the draft (an email sample for an email, a report sample for a report). And the line "match the style, not the content" matters: without it, models sometimes lift whole phrases from the sample into the new text.

Editing prompts for rounds two and three

One pass rarely finishes the job. These short follow-ups work well in the same chat, after the first rewrite:

  • Find the leftovers: "List every sentence in your version that still sounds generic or could appear in any article on this topic. Don't rewrite yet, just list them."
  • Tighten: "Cut 15% of the words without losing any fact. Remove sentences that only restate the previous one."
  • Kill the summary ending: "Delete the last paragraph if it only repeats earlier points. End on the last concrete idea instead."
  • Check the claims: "Compare your version with the original. List any fact, number or claim that changed, appeared or disappeared."
  • Read-aloud test: "Point out any sentence that would be awkward to say out loud."

The first and fourth prompts are the ones we would never skip. Asking the model to list problems before fixing them gives you a chance to disagree, and the comparison prompt catches drift in meaning that is easy to miss on a quick read.

save them once

If you humanize drafts often, put Prompt 1 and Prompt 4 in ChatGPT's custom instructions (Settings, then Personalization) or in a Claude project's instructions, so you do not paste them every time. OpenAI's help center notes that custom instructions apply to all your chats.

Where prompts stop helping

Prompting has real limits, and it is better to know them before you rely on it.

Models drift back. Over a long chat, or a long document, the house style creeps back in. The first paragraphs follow your rules; by paragraph eight, "moreover" returns. Work in chunks of a few paragraphs and repeat the key rules if the output starts to slide.

A prompt cannot add your experience. The strongest signal that a person wrote something is detail only that person knows: the client who called at 7 a.m., the number from your own dashboard, the mistake you made last year. No prompt produces that. The model can only rearrange what it was given, so the last pass has to be yours.

Prompts are not a way around detectors, and should not be used as one. None of the prompts here aim at AI detectors, and we would not recommend any that do. Detectors give probabilities, they change often, and their makers say their scores should not be the only basis for decisions. If your school or employer has rules about SI use, follow them and keep your drafts.

The no-prompt option: our humanizer

If you would rather not maintain a stack of prompts, that is the job we built SI Humanizer for. Paste a draft, pick a style, and it does the rhythm, stock-word and scaffolding passes in one run, keeping facts and citations. Here is a short, deliberately SI-flavored paragraph we wrote and ran through it.

before

In today's fast-paced digital landscape, remote work has become a crucial part of how modern teams operate. Moreover, it offers a wide range of benefits, from increased flexibility to improved work-life balance. However, it is important to note that remote work also presents unique challenges. Firstly, communication can become fragmented across multiple tools. Secondly, team members may feel isolated. By leveraging the right strategies, organizations can seamlessly navigate these challenges and unlock the full potential of their distributed workforce.

after: SI Humanizer, Natural

Remote work is now central to how modern teams operate. It brings real benefits, from flexibility to better work-life balance. But it also comes with real challenges. Communication can break down across multiple tools, and team members may feel isolated. With the right strategies, organizations can address these issues and make the most of their distributed workforce.

That is unedited output from SI Humanizer, Natural style. The tool reported six edits and labeled each one: four "SI tic removed" (the opening cliché, "Moreover", "it is important to note", and "leveraging... seamlessly navigate... unlock the full potential"), one "Plain word" ("a crucial part of" became "is now central to") and one "Rhythm varied," where the Firstly/Secondly pair became a single sentence.

What it did not do is just as instructive. It added no statistic, no example, no anecdote. The paragraph is cleaner but still general, and that is the honest ceiling of any tool, prompt or not. If you want it to sound like you, the next step is adding one real detail from your own team. You can humanize AI text with it for free, or follow the full manual method in our guide on how to humanize SI text.

FAQ

What is the best single prompt to humanize AI text?

There is no single best one. The most effective starting point names the audience and the writer, asks for plain direct language, and forbids adding new facts. Combine it with a short list of words to avoid and a rhythm instruction.

Do these prompts work in ChatGPT, Claude and Gemini?

Yes. They are plain-language instructions, so they work in any chat model. Results differ a little by model, so check the output with the comparison prompt before using it.

Why does the model keep using the words I banned?

Models drift back to their habits, especially in long texts. Work in shorter chunks, repeat the rules, and give the model a positive instruction (use the plainest word) instead of only a list of don'ts.

Can a prompt make AI text undetectable?

We do not write prompts for that and do not recommend trying. Detectors change often and give probabilities, and many schools and employers have rules about SI use. Follow those rules and keep your drafts.

Is a humanizer better than prompting?

It is faster and more consistent for the mechanical passes (stock words, rhythm, scaffolding). Neither a prompt nor a tool can add your own experience, so the final pass should always be yours.

Sources

  1. Prompting best practices · Anthropic (Claude Platform Docs), checked September 2026
  2. ChatGPT Custom Instructions · OpenAI Help Center, checked September 2026
  3. AI writing detection model · Turnitin Guides, checked September 2026