In short
Edit AI writing in two passes: draft first, then run a separate cleanup that cuts the patterns readers spot, like binary contrasts, throat-clearing openers, recap endings and clusters of em dashes [1]. Then add what only you have: a detail from your week, sentence lengths that follow your speech, and an ask in your own words. Give the AI your own posts and a few rules before the next draft, read the result aloud, and keep an edit only if it still sounds like you [1].
Why AI drafts sound alike
You ask ChatGPT for a post about your work and get something clean, correct and easy to scroll past. AI writes clean drafts fast, and clean AI drafts all sound alike [1]. The model reaches for the most expected sentence every time, so ten coaches asking for a post on burnout get ten posts with the same bones. Your clients have seen those bones many times over. They spot "Here's what nobody tells you" in one line and keep scrolling [1].
If your drafts sound like everyone else's, look at what went in before you blame your writing. The AI never sat in your client meetings, so it fills the gap with the average of other people's posts. Wikipedia's editors, who track AI writing closely, describe the same thing: the model leaves out specific, unusual facts and puts more generic, positive descriptions in their place [2]. Every new chat starts from zero as well, so you re-explain your tone each time [3]. Give the AI your material before it drafts, and take its habits out in a second pass.
This hits hardest when you sell your own expertise. A coach, consultant, real estate agent, freelancer or lawyer gets hired for how they think, and a post that sounds like everyone else's makes you look like every other option.
The patterns to cut
These are the patterns that give a draft away [1]. Look for them in your own drafts before you look anywhere else:
| Looks like | Fix | |
|---|---|---|
| Binary contrast[1] | "It's not X. It's Y." | Say what it is, and drop the straw man |
| Throat-clearing opener[1] | "Here's the thing", "Let me be clear" | Start with the point |
| Faux-insight setup[1] | "What nobody tells you", "The part everyone misses" | Say the thing itself |
| Colon reveal[1] | "The best part: it learns." | Write it as a plain sentence |
| Superficial analysis[1] | A sentence that trails off "highlighting" or "underscoring" something | Cut the trailing clause |
| Importance puffery[1] | "a testament to", "marks a pivotal moment" | Say what changed |
| Weasel attribution[1] | "Experts agree", "studies show" | Name the source, or cut the claim |
| Synonym cycling[1] | "the system", then "the framework", then "the method" | Pick one word and keep it |
| Negative listing[1] | "No fluff. No filler. No hype." | Say what's there instead |
| Dramatic fragment[1] | "And then? Nothing." | Join it to the sentence it belongs to |
| Fake-deep ending[1] | A one-line moral at the bottom of the post | End on the last useful thing |
| Recap ending[1] | A closing paragraph that restates the post | Delete it |
| Rule of three[2] | Three adjectives or three short phrases in a row, every time | Keep the one that's true |
| Formatting slop[1] | Emoji headers, random bold, bullets where two sentences would do, em dash clusters | Write it as plain sentences |
Some of these can hide a good line. A contrarian first line can look like a binary contrast, so if your hook works, protect it [1].
Edits only you can make
The cleanup pass can't add what the AI never had. Each of these edits needs something from your week, so make them by hand afterwards.
Swap each abstraction for something you saw. Phrases like "financial challenges", "client success" and "gaining traction" could sit in anyone's post. Put the thing itself in their place: the client's words, the document, the day it happened. Before a draft, write down one fact only you could supply, like a number you can back, a client's question in their words, or a job that went wrong.
Example
A real estate agent's AI draft: "Many homeowners underestimate the importance of preparing their property for sale." With one fact from the week: "Three sellers asked me this month why their flat got no viewings. All three had listing photos taken at night."
Let the sentence lengths follow your speech. Read the paragraph out loud, and where you'd stop, put a full stop. Where you'd run two thoughts together, let the sentence run on.
Example
A coach's AI draft: "Setting boundaries at work is important. It helps you protect your energy. It also improves your relationships." Read aloud and rewritten: "Most of my clients don't need a better morning routine. They need to stop answering work email in bed, which is harder than it sounds when the boss sends them at eleven."
Keep the asides you'd really say. Spoken advice has small interruptions, like "to be fair", "which surprised me" or a question back to the reader. Add one where you'd say it to a client, and skip any you wouldn't, because a borrowed "Honestly?" sounds more fake than none.
Example
A freelance designer's AI draft: "A clear brief leads to better design outcomes." With the asides they'd say: "Send me the brief you'd send a friend. Really. If they'd understand it, so will I, and so will whoever works on it after me."
Bridge into a list with a true story, one or two sentences about when you learned the steps. Clients remember stories and ask about them later, so never invent one.
Example
A bookkeeper's AI draft: "Here are three steps to prepare for an audit." With a bridge: "A client called me on a Friday with an audit letter and a shoebox of receipts. This is the order we sorted them in."
Write the ask the way you'd say it across a desk, with one thing for the reader to do.
Example
A lawyer's AI draft: "Contact us today to learn more about our services." The version they'd say: "If your lease renews this year, comment LEASE and I'll send you the clause checklist."
Use plain verbs. Wikipedia's editors found plain "is" and "has" phrases turn up more in human writing than in AI text [2], so when a draft says your service "offers" or "features" something, try "is" or "has" first.
Then delete any sentence a reader wouldn't miss. If the post lost no information, leave it deleted.
The five-minute check
Run this on every draft before it goes out. Every box is something you can fix on the spot.
[ ] The first line names my reader's problem, in words they'd use [ ] There's at least one fact only I could know [ ] No phrase that could sit in anyone else's post [ ] The sentence lengths sound like me reading it aloud [ ] Any aside is one I'd really say [ ] Any story is true, and I'd be happy to be asked about it [ ] The ask is one thing, said the way I'd say it to a client [ ] None of the patterns from the table [ ] I read it aloud, start to finish, and fixed what tripped me up
If the second box stays empty, stop editing, because polish can't add your material. Write down one question a client asked you this week and ask for a new draft built on it. And if a draft still reads like a machine after all this, record yourself saying the post to a client, transcribe it, and edit the transcript instead.
Draft first, clean second
Write the draft however you like: with ChatGPT, with Claude, with a writing prompt you trust, or by hand. Then run the cleanup as a separate step [1].
The split matters because the formulas that make drafts punchy are the same ones that leave the tells. Draft and clean in one prompt and you get a half-clean draft. Two passes keep the structure and lose the tells [1].
Use this as your cleanup pass. It makes the smallest edit that fixes the draft [1], so your own words and your humour stay put:
Edit the draft below. Make the smallest changes that fix it, and keep my vocabulary, my humour, my bluntness and any honest uncertainty.
Cut these patterns:
- binary contrasts ("It's not X. It's Y.")
- throat-clearing openers ("Here's the thing")
- faux-insight setups ("What nobody tells you")
- colon reveals ("The best part: ...")
- sentences that trail off "highlighting", "underscoring" or "showcasing" something
- importance puffery and weasel attribution ("experts agree", "studies show")
- synonym cycling, negative lists, dramatic fragments, fake-profound endings and recap endings
- emoji headers, random bold, bullets where two sentences would do, and em dashes
If a claim needs a number or a name I haven't given you, ask me for it. Never invent a statistic or a source.
Keep the first line exactly as written.
Then list what you changed and why.
Audience: [who it's for]
Goal: [what they should do next]
[paste draft]Read the list of changes, and undo any change that made the post sound less like you [1]. If you keep your own writing prompts, such as a saved post prompt in a Claude project or a custom GPT, add the edit pass as their last step [1]:
FINAL STEP When I approve the draft, run the edit pass on it. Keep the first line exactly as written. Show me the list of what you changed.
Keep your drafting prompt for structure and let the edit pass do the cleanup. Each one stays short, and when a draft goes wrong you can tell which step did it [1].
Skip the cleanup for satire or deliberate hype, for short DMs with a fixed format, and for legal or medical wording that has to stay exactly as approved [1]. A three-line DM has little room for slop anyway [1].
Audit a draft before you edit
Sometimes you want the list without the rewrite: a draft from a colleague, a ghostwriter or a freelancer. Ask for findings only. Each finding points to a line you can change, which is more use than a detector's score [1].
Don't rewrite anything. Read the draft below and list every AI writing pattern you find. For each one, give me: - the pattern's name - the exact line, quoted - a short fix No score, and no guess about whether AI wrote it. [paste draft]
Run it on your own swipe file too. Paste a post that did well as Post A and one that flopped as Post B, and compare the two lists. If the flop has more of the patterns, treat that as a hunch to test on your next few posts before you call it a rule [1].
Try it on one post today
Start with a post you've already published [1]:
- Run the detect pass and read every finding.
- Run the edit pass on the same post.
- Make the edits only you can make, then run the five-minute check.
- Read both versions aloud, and keep the edit only if it still sounds like you [1].
Here's what that looks like on one LinkedIn post by an employment lawyer:
Example
The AI draft: "Employment contracts aren't just paperwork. They're your protection. Many employees overlook important clauses, which can lead to problems down the line. Understanding your contract is essential for safeguarding your career. Always read before you sign!"
Example
After the cleanup and the edits: "Three people asked me this month whether they can join a competitor after they resign. All three had signed a twelve-month non-compete they'd never read. Before you sign, find the restrictive covenants clause and check how long it lasts and how far it reaches. If you'd like me to look at yours, message me CONTRACT."
- The binary contrast about paperwork and protection went, so the post starts with the point.
- The vague clauses and problems became one named clause and what it does.
- The fact only the lawyer had, the same question from three people in a month, moved up to the first line.
- The slogan at the end became one ask, in words the lawyer would use on the phone.
Aim for a post a client would believe you wrote, and rewrite any line you wouldn't say to them across a table in the words you'd actually use. Treat the AI's version as a first draft every time. It gives you the structure, and the facts, the stories and the ask come from you.
Each draft in the Postmosis post writer shows a sounds-human check before you post, next to a predicted engagement score and the character count for every network.
What humanizer tools actually do
A humanizer such as Humanize.io or Wordtune paraphrases your draft. You paste the text in and wait [4], or highlight the part you want reworded and pick the suggestion that sounds most like you [5]. One writer who tried Humanize.io found the meaning stayed the same while the tone and rhythm changed [4]. That's why the result reads smoother: it breaks up the even beat and swaps the stiffest words.
What it can't do is add the client question, the number or the story, because your draft is all it has. So a humanizer can loosen a paragraph you already like, and it can't make a generic post yours. Use one after your own edits, if at all.
Fixing it by hand runs into the same wall if you only swap words. Change "utilise" to "use" and the sentence keeps the AI's shape, so you fix one paragraph and the next still sounds off. The same writer found that some parts got better but not across the board [4]. Rewrite each sentence from what you'd say out loud instead, then run the five-minute check.
Match the tone to the place
Your voice stays the same everywhere, but the tone shifts with the place, the way you'd talk differently to a new lead and to a client of five years. Tell the AI where the words are going:
| How it should sound | Tell the AI | |
|---|---|---|
| LinkedIn post | Like explaining something to a peer at an industry event | Plain words, one idea, keep my first line |
| Instagram caption | Like talking to a client you already know | Shorter and warmer, and the first line repeats what's on screen |
| Email to your list | Like a note to one client | Write to one person, one link, no headings |
| Reply to a lead's message | Like a text | Three lines at most, answer their question, ask one back |
| Proposal or client email | Plain and exact | Keep every number and date I gave you, and add no warmth I didn't write |
Rewrite this for [LinkedIn / Instagram / an email to my list / a reply to a message / a proposal]. Make it [more casual / more formal / shorter]. Keep my words wherever they still fit. Keep every fact, number and name. Don't add phrases I didn't use, and don't add enthusiasm. Then list what you changed. [paste draft]
A freelance designer whose proposal reads "Super excited to jump in on the rebrand!" can ask for more formal and get "I'm looking forward to starting the rebrand", with the Friday deadline and every other fact left alone.
Give the AI your voice
Every chat starts from zero. You paste context, explain your tone, correct the output, and do it all again next time [3]. Two files and one habit fix that. Run them as a loop of four steps:
- Keep a facts file with what you sell, who it's for, the questions clients ask and the proof you can back. Every draft pulls from it [3].
- Set your voice rules once: tone, banned words, who you write for and how you lay a post out [3]. Save them as project instructions in Claude or custom instructions in ChatGPT.
- Draft from both files, every time, so the AI has your facts and stops filling gaps with generic claims. The evidence file in the LinkedIn content strategy is a good template.
- Review each draft and turn what you learn into a rule [3].
Build the voice rules from your own posts. Paste three to five you're happy with and ask the AI to name the patterns, then correct its summary, because it will get some of it wrong [6]. No posts of your own yet? Record a voice note answering a question a client asked this week, transcribe it, and use that as your first sample.
Here are [3 to 5] posts I wrote myself: [paste] Describe how I write, under these headings: - Tone: how formal, how funny, how direct - Topics and angles I keep coming back to - Format: length, structure, line breaks, headings - Vocabulary: words and phrases I use, and words I never use - Style: how I open, how I close, how I ask for something Quote a line from my posts as evidence for each point. Don't add anything you can't point to.
Then turn the summary into rules you keep, under the same five headings [3]:
| Write down | Example | |
|---|---|---|
| Tone[3] | How formal, how funny, how direct | Direct, a little dry, never salesy |
| Topics[3] | Subjects and angles you return to | Always start from a client question |
| Format[3] | Length, structure, headings | Short paragraphs, no headings in posts |
| Vocabulary[3] | Banned words and preferred terms | Say "clients", never "customers" |
| Style[3] | How you open and close, and how you ask | End on a question, with no sign-off |
Keep the file alive. Every approval, edit and rejection has a rule in it [3]. Approve a draft and note what to keep. Edit one and look at what you changed, because the gap between its draft and yours is your voice [3]. Reject one and write the reason down as a "never" rule [3].
In Postmosis, Your voice reads 3 to 5 of your past posts and writes a voice summary that every draft in Write uses.
Where the cleanup pays off
Give the cleanup the audience and the goal, so it knows what to keep. These four places get the most out of it:
| What to tell it | |
|---|---|
| Every LinkedIn post[1] | Who the post is for, and to keep the first line as written |
| Newsletters, blog posts and video scripts[1] | The audience, the goal, and the asides or story to keep |
| Sales pages and cold email[1] | The reader, the word limit, and to ask you for any missing fact |
| Your own writing prompts[1] | Add it as the last step: run the edit pass and show what changed |
On sales copy, the edit swaps puffery and weasel claims for specifics from your draft, or asks you for the missing fact [1]. When it asks for a number or a client name, give it a real one or cut the claim.
Prefer a ready-made tool? No AI Slop is a free, MIT-licensed skill by Peter Yang that runs the same kind of edit, without inventing stats or sources, inside Claude Code, Codex or any agent that installs skills [1]. It's also listed as a plugin in ChatGPT [1].
Example
Instructions for a cold email: "Cold email to the operations lead at a 40-person logistics firm. Under 120 words. Keep my opening question." Then paste the email underneath.
What about AI detectors
The readers who matter are already detecting. LinkedIn says posts that feel generic, repetitive or lack a clear point of view are less likely to be widely distributed [7], and members can mark a post with "Seems like AI slop" [7]. If enough of them do, you'll see a tip in your post analytics [7].
Don't write for a detector. OpenAI pulled its own AI text classifier on July 20, 2023, because of its low rate of accuracy [8]. In OpenAI's tests it caught 26% of AI-written text and wrongly flagged human writing 9% of the time [8], and it was very unreliable below 1,000 characters [8].
AI text can also be edited to slip past a classifier [8], so a score tells you about the surface and little else. The person reading is the one who decides whether to call you.
If you leaned heavily on AI for a LinkedIn post, LinkedIn recommends saying so when it isn't obvious from the context [7].
Words to ban outright
Patterns matter more, but a short ban list catches the laziest drafts. Wikipedia's editors treat a post-2022 edit packed with words like these as one of the strongest tells of AI use [2]. Paste this into your voice file or straight into your prompt:
Never use these words: delve, tapestry, testament, pivotal, crucial, robust, harness, intricate, interplay, foster, garner, showcase, vibrant, boasts, underscore, enhance, meticulous, realm, landscape (as an idea), journey (unless it's a trip), navigate (unless it's a map). Never use these phrases: "in today's fast-paced world", "it's important to note", "at the end of the day", "in conclusion", "let's dive in", "I hope this helps". Cut empty adverbs (really, truly, incredibly, actually) unless that's how I talk. No em dashes. Use a full stop or a comma.
Keep the list short. A long one makes your writing stiff in a new way, and the AI starts swapping one stock word for another.
Word lists also go stale. Wikipedia's editors note that ChatGPT overused "delve" in 2023 and that it dropped off sharply in 2025 [2]. The words change, but the patterns underneath them stay, which is why the patterns come first.
Sources
- Remove AI Slop from Your Writing, Claude Master Guide, Keepy Labs, read 29 October 2026
- Signs of AI writing, Wikipedia, read 29 October 2026
- How AI CMO Learns Your Brand Voice, Keepy Labs, read 29 October 2026
- Struggling with AI writing that feels robotic?, Rubin on Medium, read 29 October 2026
- Humanize AI, Wordtune, read 29 October 2026
- LinkedIn Content System, Claude Master Guide, Keepy Labs, read 29 October 2026
- Best practices for content created with the help of AI, LinkedIn Help, read 29 October 2026
- New AI classifier for indicating AI-written text, OpenAI, read 29 October 2026
