How to make ChatGPT write like a human (and Claude too).

1. Why ChatGPT and Claude drafts read as AI.

A language model writes the most likely continuation of what you gave it. Give it one line (“write a blog post about publishing frequency”) and the most likely article is the average of every article on that topic, written in the average voice. That voice has habits, and after three years of reading model output, people have learned them. A draft can be accurate and still lose the reader in the first line because it sounds like every other draft.

The habits are documented. WikiProject AI Cleanup, a group of volunteer Wikipedia editors who have been removing AI-written text from articles since late 2023, keeps an advice page called Signs of AI writing. It catalogues the patterns they keep finding: overuse of the em dash, “not just X, but also Y” contrasts, a cluster of favourite words (additionally, delve, pivotal, testament, tapestry), sentences that end on “highlighting the importance of”, claims credited to unnamed experts, and bold text everywhere. Section two turns the most common ones into a checklist with a fix for each.

The same page carries a warning worth taking seriously: the signs are descriptive, not prescriptive, and the editors ask people not to treat them as the problems to be fixed, because that only makes detection harder. Read that as a writer and it is good advice. The tells are symptoms of a draft with nothing specific to say. Strip them and you get clean, empty prose (version two of the worked example below). The fix that lasts starts upstream, with a brief that carries real facts, which is the process in our brief-first guide to ChatGPT for content creation. This page covers the other half: the instructions and the edit that remove the patterns.

Newer models produce fewer tells by default. Claude Opus 5.5's launch post says it communicates more naturally, and ChatGPT's August 2026 update to GPT-5.6 Sol promised more direct answers with tighter formatting. The patterns still show up in long drafts and on thin briefs, on every model, which is why the rest of this guide is model-agnostic. If you are still choosing the model itself, the comparison is on Claude vs ChatGPT for writing.

2. The signs of AI writing, with a before and after for each.

These are the eight patterns that show up most in ChatGPT and Claude drafts for blogs and marketing pages. Each one is harmless once. The problem is frequency: three or four of them in a paragraph is what a reader recognises before they can say why. Every entry has the tell, a before, an after, and the rule that prevents it, ready to copy into the instructions in sections three and four.

1. TELL 01

   Dash-spliced clauses

   Two clauses joined by a long dash (the em dash), often several in one paragraph, sometimes in pairs that bracket an aside. It is the most-quoted tell of all, and Wikipedia's Signs of AI writing page lists overuse of it among the style signs. We do not print the character on this page; [long dash] marks where the model put one.

   Before

   Most teams start a blog with real energy [long dash] and then the schedule slips [long dash] usually by the third week.

   After

   Most teams start a blog with real energy. By the third week, the schedule slips.

   The rule: Join clauses with a period, comma or colon. Put an aside in parentheses or make it its own sentence.

2. TELL 02

   Stock openers

   A first sentence that could open any article on any subject: the scene-setting line about a fast-changing world, "when it comes to", or the question restated as a statement. Readers skip it, and so do the answer boxes that quote the first sentence that actually answers something.

   Before

   In today's fast-paced digital landscape, content marketing has become more important than ever.

   After

   A blog with no search traffic usually has one problem: it answers questions nobody types into Google.

   The rule: Open with the answer or the most specific fact you have.

3. TELL 03

   "Not just X, but Y"

   A contrast that corrects a claim nobody made, so a plain point sounds like an insight. Wikipedia files it under negative parallelisms, together with "It's not X, it's Y". One per article reads fine. One per section is a signature.

   Before

   SEO is not just about keywords, it's about understanding your audience.

   After

   Pick the keywords your buyers type, then write the page that answers them better than what ranks today.

   The rule: State the point as a plain claim. Leave out the view you are arguing against unless someone actually holds it.

4. TELL 04

   Moreover, furthermore, additionally

   Formal connectors that glue paragraphs together without adding meaning. People rarely use them in a blog post. Models open every second paragraph with one, and "additionally" is on Wikipedia's list of AI vocabulary.

   Before

   Moreover, internal links help search engines understand your site. Furthermore, they keep readers on the page longer.

   After

   Internal links show Google which pages belong together, and they give a reader somewhere to go next.

   The rule: Start paragraphs with their subject. Use and, but or so, or no connector at all.

5. TELL 05

   Stitched-in quotes

   Short phrases in quotation marks dropped into a sentence and credited to nobody in particular ("experts", "industry leaders"), so the paragraph sounds sourced without naming a source. Wikipedia calls the credit part vague attribution.

   Before

   Experts call long-form content "a cornerstone of modern SEO" that is "reshaping how brands compete".

   After

   Google Search Central's guidance says it focuses on the quality of content rather than how it was produced.

   The rule: Quote full sentences from a named source, or paraphrase and name the source in the sentence.

6. TELL 06

   Lists of three

   Adjectives, benefits and examples arriving in threes, sentence after sentence, because three items sound complete. One triple is normal writing. Four in a paragraph is a pattern a reader feels before they can name it.

   Before

   Our process is fast, reliable and scalable, helping teams save time, reduce costs and grow faster.

   After

   The process takes about two hours per article, and most of that is the fact check.

   The rule: Use the number of items you actually have. If there are three, give one of them a specific detail.

7. TELL 07

   Repeated phrases

   The same word or sentence shape coming back within a paragraph or across sections: the keyword in every sentence, every section ending on a one-line summary, the same verb three times in a row.

   Before

   Consistency is key to SEO success. A consistent schedule builds trust. In short, consistency is what separates winners from the rest.

   After

   Publish on the same days each week, and plan a month of topics ahead so a busy week does not break the run.

   The rule: Make each point once. End a section on its last point, never on a recap of it.

8. TELL 08

   Inflated words and -ing tails

   Vocabulary that makes everything sound significant (pivotal, testament, tapestry, delve, underscore, landscape) and sentences that end on an -ing phrase explaining why the sentence mattered. Wikipedia's page tracks both, and notes that the favourite words change from one model generation to the next, so a fixed ban list goes stale.

   Before

   This feature serves as a testament to our commitment to users, highlighting the importance of seamless onboarding.

   After

   The new setup screen asks for one thing, your site's address, and fills in the rest.

   The rule: Use plain words, and cut any trailing -ing phrase that only tells the reader the sentence was important.

Two formatting habits travel with these. Models bold a phrase in every paragraph and turn prose into bullet points, and both make a blog post look like a slide deck. Wikipedia's page lists overused boldface as a sign in its own right. The other is the closing section that restates the introduction under a heading like “Conclusion” or “Final thoughts”. End on the last useful point, or on the next step the reader takes. If the draft is headed for WordPress, the publishing side of the same problem is in our walkthrough on writing WordPress posts with AI without publishing slop.

3. How to make ChatGPT write like a human: the instructions.

Rules typed into a chat apply to that chat and fade as the conversation grows. Rules in ChatGPT's settings apply to every chat. There are two places to put them. Custom instructions live under Settings, then Personalization, and apply everywhere; since July 15, 2026 the box holds 5,000 characters on Plus, Pro, Business, Enterprise and Edu, and 1,500 on Free. Project instructions apply only to chats inside that project and take priority over your custom instructions, which makes a Project the right home for one site's writing rules. The Project setup for SEO work, briefs and link lists included, is in our ChatGPT for SEO guide.

Here is a block built from the eight rules in section two. It runs to about 1,100 characters, so it fits the Free limit too. Paste it into custom instructions, in the box that asks how you would like ChatGPT to respond, or into a Project's instructions.

Write plain, specific prose for a reader who skims and distrusts filler.
- Open with the answer or the most specific fact. Never open with the state of the world or a restated question.
- Join clauses with a period, comma or colon, and put asides in parentheses. Do not use em dashes or en dashes.
- Make each point directly. Do not write "not just X, but Y" or "it's not X, it's Y".
- Start paragraphs with their subject. Do not use moreover, furthermore, additionally, ultimately or in conclusion.
- Quote only full sentences from a source I named. Never put short phrases in quotation marks.
- Use as many list items as really exist. Do not default to three.
- Say each point once. Do not end a section by summarising it.
- Use plain words. Avoid delve, pivotal, testament, tapestry, landscape, leverage, robust, seamless and elevate.
- Do not end a sentence with an -ing phrase about why it matters.
- Write paragraphs. Use bullets only for real lists, and bold only for terms the reader must not miss.
- Mix short sentences with long ones.
- If you do not know a fact, say so. Never guess a number.

The dash rule used to be the one ChatGPT ignored. That changed in November 2025, when Sam Altman said that if you tell ChatGPT not to use em dashes in your custom instructions, it now does what it is supposed to do. The rule is phrased with the alternative (period, comma, colon, parentheses) because a bare ban leaves the model guessing what to use instead, and the guess can be a semicolon in every sentence.

Test the block before you trust it. Run the same one-line request in a new chat before you save the instructions and again after, on the same model, then count the tells in each reply against section two. Expect most of them gone and a few back on long outputs: instructions tend to slip as a draft gets longer, which is why drafting one section at a time beats asking for 2,000 words in one go.

Which model you are on changes how often the patterns come back, not whether they do. As of October 4, 2026, the ChatGPT chat window runs GPT-5.6 Luna on the Free and Go plans and GPT-5.6 Sol on Plus and Pro (Pro, Business and Enterprise also get GPT-6 Astra in chat, labelled GPT-6 Pro). GPT-6.1 Sol, released September 29, runs in ChatGPT Work, Codex and the API but not yet in the chat window. Which of those to write with, job by job, is in our guide to the best ChatGPT model for writing.

4. How to make Claude write like a human.

The pattern list is the same. What changes is how Claude reads instructions, and Claude's prompting docs are specific about it. Three points from their best practices page matter for writing. Tell Claude what to do instead of what not to do: “write in flowing prose paragraphs” steers better than “do not use markdown”. Explain why a rule exists, because context about your goal helps it apply the rule to cases you did not list. And match the style of your prompt to the output you want: the docs note that removing markdown from your prompt can reduce the markdown in the reply, so a rule list written as bullets invites bullets back.

Here is the same rule set rewritten that way, as plain paragraphs inside tags, with the reason stated up front and a slot for a sample of your writing. Put it in a Project's instructions (Projects are on every Claude plan, Free included) so every chat in the project starts from it.

<writing_rules>
These articles are read by founders who skim and stop at the first sentence that says nothing, so every sentence has to carry information.

Write in plain paragraphs that start with their subject. Join clauses with periods, commas or colons, and put asides in parentheses.

Open each piece with the answer or the most specific fact available.

Make each point once, directly, as a statement. Contrast framing such as "not just X, but Y" reads as filler to this audience.

Use a list only for discrete items, with as many items as really exist.

Quote only full sentences from the sources named in the brief, and name the source in the sentence.

Prefer plain words: "use" over "leverage", "important" over "pivotal", "shows" over "underscores".

Vary sentence length. A four-word sentence after a long one is fine.
</writing_rules>

<voice_sample>
[Paste 500 to 1,000 words of your own finished writing here.]
</voice_sample>

The voice sample does more work than any rule. Claude's docs call examples one of the most reliable ways to steer its format, tone and structure, and recommend three to five of them. For writing, that means 500 to 1,000 words of your own finished prose, ideally two or three different pieces. Claude also has a built-in place for this: Styles. Next to the presets (Normal, Concise, Explanatory, Formal), you can create a custom style by pasting or uploading a writing sample, and pick it from the style menu in the chat box. The full setup, Project plus style plus staged prompts, is in our guide to setting Claude up as a writing assistant.

On models: as of October 4, 2026, Claude Opus 5.5 (released September 22) is in the model menu on the paid plans, and Claude Sonnet 5.5 (released September 28) is available to anyone, Free plan included. Opus 5.5's launch post is the most direct of any recent release about writing quality, saying early testers found its writing clearer and easier to follow. Clearer is not the same as free of tells, and long drafts on either model still need the sweep in section six. The choice between them per writing job is laid out in our guide to the best Claude model for writing.

5. Claude humanizer skills: what they fix and what they miss.

When people search for a Claude humanizer, they usually mean an agent skill: a folder of instructions that Claude Code, the Claude apps and other agents load on demand. The best known is the open-source humanizer skill, which rewrites a pasted draft to remove the signs of AI writing and builds its pattern list from the same Wikipedia page this guide draws on. It installs in Claude Code and also works in Claude.ai, Claude Desktop and other agents that support skills, such as Codex. Several forks and look-alikes exist with their own lists.

A humanizer skill is useful for exactly one job: a consistent pattern sweep that does not depend on you remembering the list. It is the section six edit pass, packaged. If you would rather own the list (your banned words, your house rules, your examples), a skill is a short markdown file you can write yourself; our guide to writing Claude skills for SEO work shows the format with copy-paste examples.

Where it stops matters more. A humanizer edits sentences. It cannot add a fact, check one, or supply the first-hand detail that makes a page worth reading, so a humanized draft is often version two of the worked example below: clean and empty. It also does nothing about claims that are wrong. And it is a different product from the “AI humanizer” tools sold to beat AI detectors, which swap synonyms and break grammar until a detector score drops. Wikipedia's own page warns that detectors have non-trivial error rates and that hiding the signs only makes detection harder. Writing for a detector is writing for the wrong reader.

6. The edit pass that catches what the prompt missed.

Instructions reduce the tells. They never remove all of them, especially past a thousand words. The edit pass is where the draft becomes publishable, and it works best as five separate steps rather than one request to “clean this up”.

1. Run the sweep as its own pass. A model checking a draft for one thing catches more than a model asked to write and police itself at once. Paste the draft into a fresh chat with the prompt below. It asks for quotes and suggested rewrites, never a full rewrite, so you see every change and the rest of the draft stays as it was.
2. Search for the words yourself. Use your editor's find bar for moreover, furthermore, additionally, “not just”, “in today's”, pivotal, testament and the dash character itself. Thirty seconds, and it catches what both passes missed.
3. Read it aloud. Rhythm problems that pass on screen are obvious out loud: five sentences of the same length in a row, or every paragraph ending on the same beat.
4. Add what the model cannot know. One first-hand detail per section: what you tried, a number from your own data, a customer example you have permission to use. This is the step that turns a clean draft into one a reader trusts.
5. Check every claim. Ask the model to list each factual claim with the source it came from, then open the sources and cut anything unsupported. An AI fact checker does this step at scale.

Below is a draft, then a list of patterns.
Quote every sentence in the draft that matches a pattern, name the pattern, and suggest one rewrite that keeps the meaning.
Do not change sentences that match no pattern. Do not add facts.

Patterns: long dashes joining clauses; stock openers; "not just X, but Y"; moreover, furthermore, additionally; short quotes without a named source; lists of three in a row; a phrase repeated within a paragraph; inflated words (pivotal, testament, tapestry, delve, underscore, landscape); sentences ending in an -ing phrase about importance.

[draft]

For a page meant to rank, the sweep is the floor, not the finish. Google Search Central says it focuses on the quality of content rather than how it was produced, and pages that read as human but say nothing new are the ones its scaled content abuse policy describes. The ranking data in our AI SEO statistics points the same way: what moves a page is the share of it a person actually added.

A worked example: one paragraph, three versions.

Take one paragraph from a post on how often a small business should publish. The firm in version three, Ledgerline, is invented, and so is everything it logged. The tells and the fixes are the ones from section two, and the dash appears as [long dash] for the reason given there.

Version two is where a prompt gets you. It is a real improvement, and on its own it would not stop a reader leaving, because it reads like advice from anyone. Version three cost about ten minutes and one conversation with the person who runs the blog. That conversation is the part of writing that cannot be prompted, and it is the part a search engine and a reader both reward.

Common mistakes when making AI writing sound human.

1. Asking it to sound human. It is the first thing everyone tries, because it is how you would brief a person. Claude's prompting guide for Opus 5.5 makes the point about design work: a general instruction to avoid a generic AI look mostly swaps one default for another, while naming specific patterns works. Writing behaves the same. Name the patterns.
2. Banning without a replacement. “Never use dashes” on its own can produce semicolons, or sentences that run on with commas. Every rule in sections three and four says what to do instead, which is the part that changes the output.
3. Running drafts through a detector-beating humanizer. A low detector score feels like proof. The rewrites that get there (odd synonyms, broken grammar, inserted typos) cost you the reader, and the detectors themselves are unreliable enough that Wikipedia's editors warn against leaning on them.
4. Fixing the words and keeping the empty structure. A draft with no banned words can still have six sections that each restate the title and a closing section that repeats the introduction. Cut the sections that add nothing before you polish sentences.
5. Overcorrecting. Deleting every list of three, every however and every long sentence produces a different kind of odd. The tells are about frequency. One triple in a page is just writing.
6. Skipping the facts because it sounds right. Human-sounding prose earns more trust, so a wrong figure in it does more damage. Check the claims last, on the final draft, every time. At ten articles a week that check is the bottleneck, which is the argument for autoblogging with a quality gate in front of publish rather than without one.