LinkedIn content prompts — 4 prompts, updated 18 September 2026

LinkedIn post prompts, and what gives an AI-written post away

Every prompt here takes material you already have and refuses to take a topic. That one rule is what separates a post somebody wrote from a post that reads as generated, and it is the reason the post writer will tell you there is not enough here to publish rather than filling the gap with a number nobody can source. Below: the 11 markers these prompts forbid and the line of each prompt that forbids it, then all 4 in order — what each needs before it runs, the assistant it is written for, and the run that goes wrong.

✓ Free, no signup, nothing to install✓ Nothing here connects to your LinkedIn account✓ All 4 in full, free, on their own pages✓ Each one names the case for not running it
See all 4, in orderWhat LinkedIn did about AI slop
4 prompts, in sequence
11 markers they forbid
9 inputs across all four
4 context files they read
The objection, before the prompts

LinkedIn stopped writing posts for people and started asking people to report them

Anyone handed a LinkedIn post prompt in 2026 has already read the backlash, so it is worth being exact about what the platform has actually done rather than what it is said to have done. Three things, all read from LinkedIn on 20 September 2026. None of them is a detector, and two of them are frequently described as the opposite of what they are.

The AI post writer is gone

Withdrawn. Its help article now reads, in full: “LinkedIn's AI-powered writing tool is not available at this time. We'll be introducing new tooling to help improve post writing in the near future.” LinkedIn gives no date for the withdrawal and none for the replacement. Guides published through 2026 still describe it as a Premium benefit; it is not a benefit of anything.

What replaced it edits, it does not write

Post Proofreader reviews a draft you already wrote — shortening it, clarifying it, proposing changes you accept or reject. English drafts only, gradual rollout, and LinkedIn says only “eligible Premium members” without naming a tier. Premium Career is $39.99 a month, read on 10 September 2026.

The slop button is feedback, not a verdict

A feedback option on a post that tells LinkedIn the post reads as AI slop. Every member. LinkedIn describes it as feeding its classifiers. It is feedback, not a report of a policy violation, and LinkedIn describes no outcome for the post you mark.

An outside detector measures something else

LinkedIn publishes no AI-content label, no detection score, and no way for an author to see whether a post was classified as AI slop. The searches for a LinkedIn AI detector have no LinkedIn product behind them that we could find. A general assistant will rewrite a draft so it reads less templated, and third-party detectors will score text as machine-written. Neither tells you what LinkedIn decided. LinkedIn classifies with its own models and publishes neither the classifier nor the score, so an outside detector is measuring a different thing.

LinkedIn's claimWhat it rests onSource
More than one million members used the feedback option within the first two weeks of launch.LinkedIn newsroom, 15 September 2026. LinkedIn reporting on its own product.news.linkedin.com/topic/product-news ↗
Views of AI-slop-classified content fell 40%, which LinkedIn credits to the feedback option together with expanded automation defences, improved classifiers, profile verification and new AI tools focused on proofreading rather than rewriting.LinkedIn newsroom, 15 September 2026. LinkedIn measuring its own classifier against its own definition of slop; no outside party can check either.news.linkedin.com/topic/product-news ↗

Both figures are LinkedIn measuring its own classifier against its own definition of slop, and no outside party can check either — which is why they are printed with the attribution attached rather than as findings. The practical reading for a writer is narrower than the headlines: the platform has told you the tone is the problem, not the tool. Full records for all three features, with what each costs and what LinkedIn refuses to say about it, are on LinkedIn's own AI features; the same three sit beside the tools and the MCP servers on the content category page.

What a reader is actually recognising

11 markers, and the line in each prompt that forbids it

None of these is a style preference. Each one is something a model produces when it has been given a topic instead of evidence — the formula fills the space where the material should be. Read the third column as a check rather than a claim: every rule below is a word-for-word line from the prompt named under it, so you can open that prompt and see it in context.

The markerWhy it reads as generatedThe rule, verbatim
An opening line that only promisesThe feed shows roughly the first three lines and collapses the rest. A line spent announcing that something surprising follows has spent the visible part of the post on nothing.“Write the opening line. It must state the claim or the surprise, not tease it.”
A statistic with no ownerThe most expensive tell, because nobody can check it — including you. A model asked for a post about a topic supplies a number, since posts about topics have numbers in them.“Never invent a statistic, a customer, a result or a conversation. Every number must appear in the material I gave you.”
The named formulas“Unpopular opinion”, “Let that sink in”, “Here’s the thing”. These are the phrases that appear most often in LinkedIn posts, so they are what a model produces when nothing else constrains it.“No LinkedIn dialect. Banned: “Unpopular opinion”, “Let that sink in”, “Here’s the thing”, “I’ll say what nobody will”.”
One word per lineA layout instruction wearing the clothes of an argument. It reads as generated because it is a format, and a format is the one thing a model can supply without having anything to say.“No hook formulas: no “Unpopular opinion:”, no “I’ll say it:”, no “Let that sink in”, no one-word-per-line ladders.”
Emoji bullets and the ↳ arrowDecoration standing in for structure. A post that needs a glyph to show that one line follows from another usually does not have the line that would have shown it.“No emoji bullets, no “↳”, no engagement bait, no “comment X and I’ll send you Y”.”
Two claims in one postGenerated posts cover a subject; written posts argue one thing. Breadth is what comes back when there is a topic and no position behind it.“One claim per post. If a draft contains two, split it into two posts and say so.”
Nothing but the general caseA named customer’s fifth entity is something that happened. Digital transformation in finance is a category, and a category is what a model reaches for when it has no material.“Specific over broad. A post about one customer’s fifth entity beats a post about digital transformation in finance.”
No stated limitationSaying where a claim stops is the part that requires having thought about it. It is almost never generated, because the limitation is not in the topic — it is in the evidence, and there was none.“State the limitation. Every post that makes a claim should say where the claim stops. This is what separates a credible post from a confident one.”
A count in the first line“3 things I learned”, “5 lessons from”. The number is a promise about structure that the post then has to be padded out to keep.“Banned: counting hooks (“3 things I learned”), and any opening line whose only content is that a surprising thing follows.”
A question nobody can answer“Thoughts?” and “Agree?” asked of a reader who has been given nothing to agree or disagree with. It is a request for engagement in place of a reason for it.“End with something answerable — a real question, or the sentence that invites disagreement. Not “thoughts?”.”
A comment that only agreesThe engagement half of the same failure. Praise is what a model writes under a post it was told to engage with, and it is invisible to everyone, the author included.“Never open with praise. “Great post” and “So true” are the noise this prompt exists to avoid.”

The list is not a detector and will not survive contact with someone who has read it. A person can strip all 11 markers out of a post that still has no evidence behind it, and what they will have is a post that reads as written and says nothing — which is why the markers are the second section here and not the first. The rule underneath them is the part that does the work.

The one rule all four are built on

Material in, never a topic — and what that does not fix

Every prompt on this page takes something that exists before the prompt runs: a transcript, a thread, a number out of your own system, a disagreement you had. Handed a subject instead, they are built to stop rather than to produce. That is an unusual thing for a prompt to do and it is the whole design, so it is worth being precise about what it buys and what it does not.

The input is evidence, not a subject

{{SOURCE_MATERIAL}} is a placeholder in two of the four: Paste the raw input — the transcript, the post, the notes, the article. The other two collect it a different way — the ghostwriter refuses to draft until three questions about what actually happened are answered, and the comment prompt wants the post pasted in full rather than linked. None of the four runs on a subject alone.

“Not enough here” is a valid answer

The post writer is told in its own body that this is a frequent verdict and must name what would make the material postable. A prompt that cannot refuse is a prompt that invents, because refusing and inventing are the only two things available when there is nothing to work from.

Every number has to be traceable to your source

Not “avoid making things up” — a specific instruction that each figure appear in the material supplied. It is the difference between a preference a model weighs and a constraint it checks against something.

The chain counts honestly instead of hitting ten

The repurposing chain rates each extracted claim strong, adequate or thin, then states how many posts the source can carry. If the answer is two, it says two. That number is the output people skip, and it is the one worth having.

It does not make the post good

A true, specific, sourced post about something nobody needed to know is still a post nobody needed. The rule removes one failure — fabrication — and leaves every editorial judgement where it was, which is with you.

It cannot tell you what LinkedIn decided

LinkedIn classifies with its own models and publishes no label and no score, so no rule followed here produces a guarantee about how a post is treated. What it produces is a post whose claims you can defend when somebody asks where the number came from.

The rule costs something, and it is worth naming: describing your voice, your company and your position in the placeholders of every run is the part these prompts are worst at. That is what the 4 context files are for — BRAND-VOICE.md, COMPANY.md, PRODUCT.md and CONTENT-ENGINE.md are .md files you attach once to a Claude or ChatGPT Project, after which every request inside it already knows the answer and the prompts get shorter. All the context files sit on their own page, each with the sections you have to fill in before it is any use.

The prompts, in order

All 4, in the order the work actually happens

Not alphabetical and not by popularity: this is the order of the content workflow, where the assistant is set up before anything is written and the multiplier comes after the thing it multiplies. Each block below carries the opening of the prompt, what it needs before it runs, the assistant it is written for, the run that goes wrong, and the situation in which running it is the wrong move. The whole prompt is one click away, on the page written for it.

Before any of them: decide the claim

Founder Positioning Brief is filed under profile rather than content, and it is still step one here. In: Your last 5–10 posts and what you sell. Out: One defensible claim, three themes, and the not-list. Without it the prompts below write competent posts that argue in four directions.

Then: decide where material comes from

CONTENT-ENGINE.md names the sources and who writes them down. A posting system with a material pipeline, rather than a calendar that dies in week three. It is the step people skip, and skipping it is why the fourth week is the one where posting stops.

1. LinkedIn Ghostwriter — Project Instruction

A standing instruction you paste once into a Claude Project or ChatGPT Project, after which every post you ask for arrives in your voice and with a critique already applied.

Where it sits

Step 3 of 6 in From raw material to a posting habit. In: BRAND-VOICE.md and the claim. Out: A Project that drafts in your voice and refuses to invent facts. 10 min, once.

What you have to have first

Four things, and it will tell you on the first request which of them is missing: a voice file built from your real sentences, the company and product files, the claim you want to be known for, and the audience. Nothing else is asked for until you ask for a post.

Claude, ChatGPT

Claude or ChatGPT, and it has to go in a Project rather than a chat. This is the one resource on the page where the assistant matters, because a Project re-reads its attached files on every turn and a chat window does not. A Gemini Gem takes the instruction; file handling there is less predictable, so paste the voice file into the instruction itself if the voice is not landing.

Put it down when

If you post less than about twice a month. Writing the voice file, creating the Project and attaching three files costs more than the posts it would save, and a standing instruction nobody runs goes stale in the same drawer as the content calendar.

What a bad run looks like

A draft arrives on the first request, without the three intake questions. That means the instruction went into a chat instead of the Project’s instruction field, or the voice file is not attached — and what you are reading is LinkedIn house style with your job title on it.

How it opens
You are the ghostwriter for {{ROLE}} at {{COMPANY}}. You write their LinkedIn posts. You have one job: produce drafts in their voice that they can publish with light edits, and never put words in their mouth that they cannot stand behind.

Another 546 words of it, plus the placeholders as fields you fill in the browser and the .md download, on the LinkedIn Ghostwriter Project Instruction page. It reads BRAND-VOICE.md, COMPANY.md, PRODUCT.md and CONTENT-ENGINE.md when they are attached to the Project, and works without them.

2. LinkedIn Post Writer

A post built from material you supplied, with three alternative openings and an honest verdict when there is not enough there to publish.

Where it sits

Step 4 of 6 in From raw material to a posting habit. In: A call, a result, a mistake, an analysis. Out: A post, two alternative openings, and the pinned comment. 5 min per post.

What you have to have first

One piece of material with specifics in it: a call transcript, a support thread, a number out of your own data, an argument you have had twice this month. A subject line is not material, and the prompt is built to say so rather than to fill the gap.

Any assistant

Any. There is no instruction in it that Claude, ChatGPT and Gemini read differently, and it is not split into per-assistant versions — a role, a stated constraint and an output schema are not model-specific.

Put it down when

When all you have is a topic. The prompt will return “not enough here”, which is the correct answer rather than a fault, but it is faster to go and get the material than to run it and be told.

What a bad run looks like

A figure in the finished post that was not in the material you pasted. Read the draft back against your source every time: a fabricated number is stated exactly as confidently as a true one, and it is the claim a reader is most likely to check.

How it opens
You are a writer who turns real material into LinkedIn posts. You do not generate content about topics; you extract claims from evidence. If the evidence is thin, you say so.

Another 424 words of it, plus the placeholders as fields you fill in the browser and the .md download, on the LinkedIn Post Prompt page. It reads BRAND-VOICE.md and CONTENT-ENGINE.md when they are attached to the Project, and works without them.

3. Repurposing Chain

A set of independent posts extracted from one source, each standing on its own claim, with an honest count rather than a forced ten.

Where it sits

Step 5 of 6 in From raw material to a posting habit. In: A long piece: webinar, report, transcript. Out: The posts the source honestly supports — an honest count, not ten. 15 min per source.

What you have to have first

One long source and nothing else — a webinar transcript, a report, a recorded call, an article you already published. Step 1 reads the source; the three steps after it read step 1.

Any assistant

Any, but all four steps in one conversation. They are not four prompts to keep in a file: each one consumes the output of the one before, so step 3 run in a fresh chat has nothing to work from and writes posts from memory.

Put it down when

When step 1 says the source supports two posts and you wanted ten. That is the chain working. Running steps 2 to 4 anyway is exactly how one honest source becomes eight thin ones.

What a bad run looks like

A claims table where every row is rated strong. A forty-minute transcript does not contain seven equally evidenced arguments — a model that grades them all strong has stopped reading the source and started agreeing with you.

How it opens
Step 1 of 4 — Extraction.
You are an editor reading source material for claims that could each carry a post on their own.

Another 470 words of it, including the 3 steps that follow this one, plus the placeholders as fields you fill in the browser and the .md download, on the LinkedIn Repurposing Prompt Chain page. It reads BRAND-VOICE.md and CONTENT-ENGINE.md when they are attached to the Project, and works without them.

4. Comment Strategy

Three comments at different levels of directness, plus a verdict on whether this post is worth commenting on at all.

Where it sits

Step 6 of 6 in From raw material to a posting habit. In: Posts from the accounts your audience reads. Out: Comments that earn a reply from the author. 10 min daily.

What you have to have first

The post itself, pasted in full, plus the author’s name and role. That is the one input that is not a placeholder, and it is deliberate: no assistant can open a LinkedIn URL, so a link in place of the text produces a comment on a page the model never read.

Any assistant

Any. The constraint that matters is the character count, and all three count characters the same way.

Put it down when

On a thread already sixty comments deep. The prompt checks for that and returns a skip verdict — a late comment in a busy thread is read by the author and by nobody else, which is occasionally the point and usually not.

What a bad run looks like

A dissenting comment on a post you actually agree with. The prompt offers the option and says to take it only where the disagreement is real, because a manufactured one is obvious to the author and stays in the thread permanently.

How it opens
You write LinkedIn comments that are worth the author's reply. Agreement is not a comment.
What I know that is relevant: {{PROOF}}

Another 307 words of it, plus the placeholders as fields you fill in the browser and the .md download, on the LinkedIn Comment Prompt page. It reads BRAND-VOICE.md when they are attached to the Project, and works without them.

Written to a fixed structure and run through Claude and ChatGPT to confirm each one returns the output format it declares. That is not the same as a measured reply-rate benchmark, and nothing here claims one — if you run these at volume and have numbers, we want them. Every prompt above is free, needs no account, and downloads as a .md from its own page. The full library — prompts, context files, agent packs and workflows across every discipline — is at the prompt library, where the map also prints the jobs nothing here does.

If you were looking for a product instead

Nothing sold as a LinkedIn content tool writes your posts

Worth saying plainly, because the phrase covers two unrelated things. Of the 42 tools researched for this site's directory, the ones filed under content act on other people's posts rather than producing yours, and the heading that would cover post writing carries no tools at all.

LinkedIn Content AI: 6 tools, none of them a writer

Tools whose AI acts on LinkedIn posts — writing comments, automating likes, pulling out the people who engaged. None of them writes the posts. Two of those three actions are named directly in LinkedIn's User Agreement, which is why the directory prints how each tool reaches LinkedIn before it prints what the tool does. The 6 entries, with what each one exposes.

LinkedIn Personal Branding AI: the heading carries nothing

The forty-five tools researched are outbound and data products. Several act on other people’s posts — a comment, a like, an export of who engaged — and those are filed under LinkedIn Content AI. None of them writes your posts, holds a calendar or works on your own profile, so this heading carries no tools rather than a stretched reading of six.

Publishing is the part a chat window cannot do

A draft leaves an assistant as text and goes into the composer by copy and paste. The two routes that actually publish are an MCP server running on LinkedIn's own posting permission, and a tool that drives your signed-in session — very different arrangements with very different consequences, recorded on the MCP server directory.

Content prompt FAQ

The six asked before anyone copies a prompt

Where an answer rests on a LinkedIn figure, the figure is read through the same tables as the rest of this site rather than retyped here, so there is one place to correct when LinkedIn moves it.

Is there a LinkedIn AI detector?

Not one you can run, and not one LinkedIn publishes. The only LinkedIn surface that touches the question is a feedback option on a post — "Seems like AI slop" — which every member has and which tells LinkedIn that a post reads as generated. LinkedIn describes it as feeding its classifiers. It publishes no AI-content label, no detection score, and no way for an author to see whether a post was classified, so there is nothing for a detector to report on. Third-party detectors will score text as machine-written, but that is a different measurement: LinkedIn classifies with its own models and publishes neither the classifier nor the result. Read from LinkedIn on 20 September 2026.

Are these prompts different for Claude and for ChatGPT?

Three of the four are not. A prompt with a role, stated constraints and a declared output schema is not model-specific, and splitting one into a Claude version and a ChatGPT version that differ in the first line is how prompt sites end up with two thin pages competing for one query. The exception is the ghostwriter instruction, and the difference there is not the model — it is the object. A Project applies its instruction to every conversation inside it and re-reads its attached files on every turn; a pasted prompt is consumed once and dies with the chat. Claude Projects and ChatGPT Projects both do that. A Gemini Gem takes the instruction but handles attached files less predictably, so paste the voice file into the instruction itself if the voice is not landing.

How long should a LinkedIn post be?

The prompts here cap a post at 120 to 220 words, and the reason is structural rather than stylistic: LinkedIn collapses a post after roughly the first three lines and most readers never expand it. A 600-word post is a three-line post with 550 words of hope attached. The same arithmetic is why the opening line has to carry the claim instead of promising one, and why the comment prompt caps a comment at 300 characters — past that, a comment is collapsed too and reads as an attempt to write your own post underneath someone else’s.

What do I post when I have nothing to post about?

Nothing on this page will answer that, and that is deliberate rather than a gap waiting to be filled. An assistant asked for ten LinkedIn topics with no input returns ten it invented, with invented numbers inside them — which is the exact failure the 11 markers above describe. What the library does instead is upstream: CONTENT-ENGINE.md is a file that names where material comes from and who writes it down, so the question stops being asked in front of an empty composer. The post writer also returns a "what I left out" section with every draft, which is usually the next post already written down.

Do I need LinkedIn Premium to use AI on my posts?

No, and Premium buys less here than the guides suggest. LinkedIn had an AI post-writing tool and withdrew it: its help article now says the tool is not available at this time and that new tooling will be introduced, with no date for either. What Premium carries instead is Post Proofreader, which edits a draft you already wrote — it shortens and clarifies, English only, and LinkedIn describes it as reaching eligible members without saying which tier or how eligibility is decided. Premium Career is $39.99 a month, read from LinkedIn on 10 September 2026. Everything on this page is free and needs no account.

Will one of the LinkedIn AI tools write my posts for me?

Not one. 45 were researched for this site’s directory and 42 published, and every one of them sells to a recruiter or a seller. The 6 filed under LinkedIn Content AI act on other people’s posts — commenting, liking on a schedule, exporting the people who engaged — and two of those three are actions LinkedIn’s User Agreement names directly. The heading that would cover writing your own posts carries no tools at all, which is recorded as a gap rather than filled by rereading a sales tool as a content one. The directory says what each entry does to the account it runs on; it does not suggest running any of them.

Next questions

Four prompts are one answer to posting on LinkedIn. These are the others, and the pages the figures in the answers above were read from.

Where to next

Find the material before you open the composer

Every one of the 4 prompts above starts from something that already happened, and the post writer will tell you when there is not enough of it. That refusal is the feature — it is the only thing standing between a post and an invented number.

See the 4, in orderWhere each one sits, what it needs first, and the run that goes wrong.The files that stop you re-describing your voiceAttach once to a Project; every later request inherits them.

We use one analytics cookie to see which pages answer real questions. No advertising, no cross-site tracking, no data sold. What we collect.