How to write LinkedIn posts with AI that don't sound like AI

You can use AI to write LinkedIn posts that read like you, but only if you hand it the parts that make a post yours: your facts, your story and your opinion. The model handles wording. You handle substance.

AI-written LinkedIn posts sound like AI because the model is given a topic and fills the gaps with the most common patterns it has seen: a dramatic one-line opener, a list of three, a contrast ("it isn't about X, it's about Y"), a moral, and a question at the end. Fix the input (facts, voice samples, banned phrases, a structure) and run one editing pass, and most of those patterns disappear.

This post gives you a copyable banned-phrase list, five structures that read like a person, a prompt, a 10-point edit checklist and two before/after examples we wrote for this article.

Why do AI-written LinkedIn posts sound like AI?

They sound like AI because they share the same shapes and the same vocabulary, and readers now recognize both within two lines. The volume is high: Originality.ai, which sells AI-detection software, estimated that 81.2% of 5,000 long-form LinkedIn posts it sampled in July 2026 were "likely AI" (Originality.ai, July 30, 2026). Treat the exact figure with the caution any vendor study deserves; the direction matches what Wired reported in 2024, that "over half of longer English-language posts on LinkedIn are AI-generated" (Wired, November 2024).

The tells fall into three groups:

• Shapes. One-sentence paragraphs stacked down the screen, lists of exactly three, a "plot twist" line, a moral in the last paragraph, a closing question nobody needs to answer.
• Words. Abstract verbs and nouns that could fit any industry. A post about invoicing and a post about hiring end up with the same vocabulary.
• Missing specifics. No names, no numbers, no dates, no mistakes. Nothing a reader could check or argue with.

The third group is the real problem. The first two are symptoms of a model filling space where facts should be.

What does LinkedIn say about how the feed treats posts?

LinkedIn says it uses LLMs to understand what a post is about and that it is reducing low-substance posts and engagement bait. In a March 12, 2026 update, LinkedIn wrote that it wants to "reduce repetitive, low-substance posts and engagement bait, such as 'comment to agree' prompts" and to "curb comment automation or third party tools that create fake conversations" (LinkedIn News).

Other statements from LinkedIn worth knowing:

• Its engineering team says ranking considers "what you've read, liked, commented on, returned back to, or simply scrolled past" (LinkedIn Engineering, March 2026). Scrolling past counts.
• LinkedIn's Tim Jurka wrote that LinkedIn uses LLMs "to help identify the context of a post and determine whether the content is a helpful insight, a job opportunity, or a career milestone" (LinkedIn, August 2025).
• The Professional Community Policies ask members to "create original, professional, relevant, and interesting content" and say "Don't do things to artificially increase engagement with your content" (policies).

None of this says AI text is penalized. It says posts with nothing specific to offer, and posts built to farm comments, get less reach. Generic AI posts happen to be both.

Technique 1: Give the AI facts and a story instead of a topic

The single biggest fix is the input. "Write a post about hiring" produces a generic post. "Here is what happened when we changed one step in hiring, with the numbers" produces a post only you could write.

Before you open the AI, write down five things in plain notes:

1. Who it is for. "Engineering managers at 20 to 100 person startups."
2. What happened. One specific event, with a date, a number and a name (or role).
3. What you learned. One sentence. If you need three, you have three posts.
4. What you would do differently. The mistake makes it credible.
5. What the reader should do. A concrete action, not "think about it."

Notes can be messy. The model is good at turning bullets into sentences; it cannot invent your quarter.

Technique 2: Ban the phrases that give it away

Models follow explicit lists far better than "sound natural." Keep a banned list, paste it into every prompt, and add to it whenever you spot a new tell. This is the list we use; copy it and extend it.

Banned words and phrases:
delve, tapestry, testament, game-changer, unlock, unleash, elevate,
supercharge, revolutionize, cutting-edge, seamless, robust, leverage,
harness, landscape, navigate (figurative), journey (figurative),
in today's fast-paced world, here's the thing, the result?,
let that sink in, read that again, plot twist, I'm humbled to announce,
thrilled to share, agree?, thoughts?, what do you think?

Banned structures:
- "It's not X. It's Y." and "not just X, but Y"
- lists of exactly three adjectives or nouns
- a one-line dramatic opener followed by a line break
- every paragraph one sentence long
- a moral or lesson in the final line
- a closing question asking for comments
- the long dash character; use commas, periods or parentheses
- emoji as bullet points

If your company has brand rules, put the brand's banned phrases on top. Our guide to brand voice rules for AI-written posts covers how to write those rules so agents follow them.

Technique 3: Use structures people actually write in

Posts written by people tend to follow a few plain shapes. Give the model one by name so it doesn't fall back to the default stack of one-liners.

Each one starts with a fact, not a hook. The first line of a changelog post can be "We removed the demo request form on March 3." That is more specific than any hook formula.

Technique 4: Give it your voice with samples

Paste three of your best past posts and describe your voice in two sentences: "Short sentences. Dry, a bit self-deprecating. Never uses exclamation marks." The samples do more than adjectives like "conversational" ever will.

Pick samples that performed well and that you still like. If you have none, use a few emails or Slack messages you wrote to customers. Those are often closer to your real voice than old posts.

Before and after: two examples

These are examples written for this article. The companies and numbers are invented to show the technique.

Example 1: a hiring post

Before (topic-only prompt: "write a LinkedIn post about hiring lessons"):

Hiring is broken.

In today's fast-paced world, finding the right talent is harder than ever.

Here's the thing: it's not about resumes. It's about potential.

3 lessons I learned this year:
(check emoji) Culture matters
(check emoji) Speed matters
(check emoji) Communication matters

The result? A team that truly thrives.

What do you think?

After (facts-first prompt, banned list, "mistake" structure):

We lost two senior engineer candidates in June to faster offers. Both told us the same thing: our process took 41 days.

So we cut it. The take-home test went (it was 6 hours of unpaid work), and the final round became one 90-minute call with the people they'd work with.

Our last four hires took 19 days from first call to offer. Offer acceptance went from 2 of 5 to 4 of 4.

What I'd do differently: we kept the take-home for a year because it felt rigorous. Nobody had checked whether it predicted anything. It didn't.

If you run hiring for a small team, time your last five hires from first call to offer. That number is worth knowing before you change anything.

Example 2: a product update

Before:

Big news. We're thrilled to share that our new reporting dashboard is LIVE.

This is a game-changer for finance teams everywhere. Seamless. Powerful. Intuitive.

Ready to elevate your reporting? Link in comments.

After:

Our customers export the same three reports every Monday, then rebuild them in a spreadsheet. We watched 12 of them do it on calls.

Those three reports are now one page that updates overnight: cash in, cash out, and invoices more than 30 days late.

It's live for every account today. If you used to do the Monday export, I'd like to hear whether this replaces it or misses something.

The second versions are longer on facts and shorter on adjectives. Neither needed a hook formula.

The editing pass: a 10-point checklist

Every AI draft gets one human pass before it goes out. It takes five minutes.

1. Delete the first line if it is a slogan. Start with the fact.
2. Search for the long dash and replace it.
3. Search for your banned list. One or two phrases usually survive.
4. Count lists of three. Cut one item or merge them into a sentence.
5. Merge one-sentence paragraphs into normal paragraphs where they belong together.
6. Check every number, name and date against your notes. Delete anything the AI added.
7. Remove the moral in the last line if the story already made the point.
8. Replace the closing question with a concrete action or remove it.
9. Read it aloud. Rewrite any sentence you would not say to a colleague.
10. Check length. LinkedIn allows up to 3,000 characters per post (LinkedIn Help, as of November 2026); most strong posts use far fewer.

A prompt you can copy

This works in ChatGPT, Claude or Gemini. Fill in the brackets.

Write a LinkedIn post for [audience].

Facts (use only these; do not add numbers, names or events):
- [what happened, with date and number]
- [what we learned, one sentence]
- [what we'd do differently]
- [what the reader should do]

Structure: [changelog | mistake | teardown | answer | disagreement].
Voice: [two sentences]. Match the voice of these three posts:
[paste post 1]
[paste post 2]
[paste post 3]

Rules:
- First line states a fact. No hook formula.
- Normal paragraphs of 2 to 4 sentences.
- Under 1,300 characters.
- No closing question, no hashtags unless I add them.
- Follow this banned list exactly: [paste banned list]

Give me two versions with different first lines.

If you run posts through an agent connected to your publishing tool, keep the same prompt as a saved instruction so every draft starts from it. Our guides on scheduling posts with Claude and with ChatGPT show that setup.

How Tracee handles AI drafts for LinkedIn

In Tracee, brand rules hold your tone and banned phrases. Agents and Tracey, the built-in assistant, read them before writing, and a version that uses a banned phrase is flagged and can't go out until it is fixed. Brand learning reads your website and past posts to build the voice profile the drafts start from. The Tracee Chrome extension works on LinkedIn posts in your feed: it writes your own version of a post you like, borrowing the angle and never the words, and drafts replies. Posts from agents wait in an approvals queue by default, so a person does the editing pass before anything is published. Details are on the AI agents page.

Next step

Take your next LinkedIn post idea and write the five fact notes before you open any AI tool. Paste them into the prompt above with the banned list and three past posts, then run the checklist. Compare the result with your last AI-assisted post; the difference usually shows in the first line.