The Humanizer Prompt for ChatGPT and Claude, October 2026 Edition

You write with AI, and you don’t want it to sound like it. The humanizer prompt that went around last year told AI to drop the em dash, “delve,” and “tapestry,” and it worked until the models changed. The newest Claude, ChatGPT, and Gemini dropped the em dash on their own, and the words that give them away now are ones the old prompt never mentions. This is the rebuilt version, from September and October 2026 measurements of what the three tools write far more often than people do. Paste it once, put your draft under it, and read what comes back out loud. It’s right below.

Copy the whole box. Paste it at the top of a new chat, then paste your draft under it and send.

FOLLOW THIS WRITING STYLE (October 2026 edition). It is based on what AI models write far more often than people do, measured in September and October 2026.

SHOULD:
- Use clear, simple words. Use, not utilize. Help, not facilitate. Start, not commence.
- Use contractions: it's, don't, we've, can't, that's, didn't.
- Vary the sentence length on purpose. Follow a long sentence with a short one. Let some run past 30 words and some stop at 5.
- Write in first person where the text is mine, and say what I think: "I think," "I'd say," "I like."
- Use the everyday words people use when they talk: pretty, a bit, a lot of, definitely, going to, check out.
- Use parentheses for a quick aside now and then (like this).
- Ask a real question when one belongs. Use one exclamation point if the moment earns it.
- Name something specific: a year, a month, a place, a person, a number, or a real example from the text. If the text has none, ask me for one. Never invent one.
- Use "you" and "your" to talk to the reader.
- Use a short list when there really are three or more separate things.
- Start with the point. End when the point is made.
- Keep every fact, name, and number exactly as given.

AVOID:
- Announcing what you're about to cover, and summarizing what you just said at the end.
- Framing the topic as the new answer to an old way of doing things.
- The shapes "it's not X, it's Y," "not just X but Y," and "more than a X, it's a Y." Say the one thing you mean.
- Groups of three for the sake of three.
- Telling the reader that something matters, is important, or is worth noting. Show it instead.
- Em dashes. Use a period or a comma.
- Long words where a short one works, and nouns made out of verbs: utilization, implementation, optimization.
- Stacked hedges: "may potentially," "could perhaps."
- Every sentence landing at the same length.
- Warnings, notes, and explanations about the rewrite. Give me only the text.
- Markdown, asterisks, and headers unless I ask for them.

AVOID these words and phrases (each one is used far more by AI than by people in 2026):
this matters, why this matters, matters because, matters most, just as important, rather than, rather than simply, rather than relying, not simply, is not simply, not merely, not automatically, together these, taken together, the result is, a real difference, extends beyond, adds another layer, what comes next, looking ahead, often called, in practice, dependable, steady, thoughtful, genuine, meaningful, practical, clearer, quietly, reliable, lasting, distinction, manageable, broader, straightforward, useful, central, furthermore, ultimately, incredibly, massive, profound, immense, highly, deeply, heavily, crucial, fundamentally, comprehensive, navigating, testament, remarkably, whether you are, we will explore, the world of, the power of, delve, landscape, nuance, foster, realm, tapestry, game-changer, unlock, elevate, leverage, seamless, robust, streamline, harness, in today's world, at the end of the day, in conclusion.

IMPORTANT: Before you answer, read your draft once more and remove any em dash and any phrase from the list above. Then list, in one line each, any sentence you weren't sure about, so I can decide.

Every line in the prompt comes from one of three studies. Graphite, a marketing research firm, compared 10,000 articles written before ChatGPT existed with 90,000 AI versions of the same topics, then added Claude Opus 5.5 on October 1, 2026. The Economist ran 1.2 million words through ChatGPT, Claude, Gemini, and Grok and compared them with its own writers. A third study measured the shape of 2,250 company blog posts against 11,250 AI copies. The prompt does three jobs, and every line in it belongs to one of them: it puts back the things people write and AI leaves out (contractions, an aside, a real date, a short sentence after a long one), it strips out the words that replaced the em dash (“this matters,” “not simply,” “rather than,” and 71 more), and it breaks the shape detectors read (the preview up top, the recap at the end, the staircase sentence). The table below shows where each line came from. Read each row left to right: what the researchers measured, then the exact line in the prompt that answers it, then what changes in your draft.

So when a draft comes back and it reads like you, that’s not luck. The contractions, the aside, the short sentence, and the real date are the human marks from the first half of the table. The missing “this matters” and “rather than” are the second half. If something still sounds off, Part 4 has the follow-ups that fix one sentence at a time. And if you want all 20 signs with the numbers for each tool, so you can spot them in anyone else’s writing, they’re in 20 New Signs AI Wrote It, Now That the Em Dash Is Gone.

Both versions below were written by hand for this guide, so you can see the tells side by side. First, the way AI tends to say it:

Customer onboarding is not simply a checklist. It is a practical, dependable process that matters because it sets the tone for the entire relationship. Rather than relying on a single welcome email, a thoughtful sequence builds trust over time. Together, these steps make a real difference.

Count them: “not simply,” “practical,” “dependable,” “matters because,” “rather than relying,” “thoughtful,” “together these,” “a real difference,” no contractions, four sentences of nearly the same length, and nothing specific. Now the way a person says it:

We used to send one welcome email and hope. Last March we switched to three short ones over a week (the second is just a photo of the team), and the number of people who actually finished setup went from about half to almost all of them. I don’t think it’s the emails. I think it’s that someone said hello twice!

A date, a parenthesis, a number, “I think,” “it’s,” “don’t,” one exclamation point, and sentences that run 9, 38, 6, and 8 words. That’s what the prompt is built to do.

Paste the prompt, paste your draft, send. Read what comes back out loud. Then use these follow-ups, one at a time, as needed:

• “Shorter. Cut a third without losing a fact.”
• “Read it back as if you’re telling a friend across a table, and change anything you wouldn’t say out loud.”
• “Which sentence still sounds like a robot? Fix only that one.”
• “Here’s a real detail to work in: [your example, your date, your place].”
• “Make the first sentence the point. Delete the warm-up.”

If you’ll use this often. Every one of the three tools has a place to save standing instructions so you don’t paste them each time: ChatGPT calls it custom instructions, Claude puts it in a Project’s instructions, and Gemini keeps it with the information it saves about you. Paste the prompt there once, and every chat starts with the tells already off.

The tell that it slipped. Long chats forget. If a reply comes back with “rather than” or “this matters” in it, the model has drifted off the instruction. Paste the prompt again at the top of a fresh message and send the draft with it. That resets it.

It changes the words. A September 2026 structure study caught AI 96 percent of the time even after the AI reworded its own work, because rewording leaves the shape: the preview, the recap, the same-length sentences, the missing details. The prompt attacks the shape too, but the strongest fix is the one line only you can write. Your example. Your date. Your opinion. Put one real thing in every piece and it stops reading like a template, because it isn’t one anymore.

The date on this list. Tells change with every model version. The em dash was the big one until September 2026, and the lists in this guide are as of October 2026. If a new model lands and the banned words in Part 1 stop showing up in AI drafts, the words have moved again, and the human habits in Part 2 are what still hold.

• Graphite, AI Tells: Opus 5.5 Update (October 1, 2026), for the Claude Opus 5.5 em dash count, the 2,548 tells, and the “this matters” and “why _ matters” multipliers.
• Graphite, AI Tells (September 16, 2026), for the method, the ChatGPT-6 Astra and Gemini 3.1 Pro tells, and the human tells. Their measurements are published under CC BY 4.0; the banned list and the table in Part 2 are drawn from their downloadable data.
• SlopShape: Identifying AI-Generated Commercial Web Content (September 14, 2026), for the structure findings and the 96.1 detection figure after rewording.
• The Economist’s analysis, as reported August 4, 2026, for the 1.2-million-word comparison, the punctuation findings, and the long-word and nominalization findings.