ChatGPT prompts for LinkedIn posts that do not read like ChatGPT

The best ChatGPT prompt for a LinkedIn post is one that stops asking the model to invent expertise it does not have. Prompts that begin “write a LinkedIn post about leadership” return the average of everything ever written about leadership, which is precisely the content our data shows does worst. The prompts below do something different: they extract material you already own, then structure it. All of them are shaped around what separated the top 10% of 12,988 ranked posts from the bottom half.

Why most ChatGPT LinkedIn prompts produce unusable posts

A language model given a topic and nothing else produces the consensus view of that topic, smoothed and hedged. That is not a flaw in the prompt, it is what the request asked for. The problem is that the consensus view is exactly what no one stops scrolling for.

Our analysis makes the gap concrete. Among 12,988 ranked posts, openers that put a specific person in a specific moment appear in 19.6% of the top decile and 10.1% of the bottom half. Posts that address the reader directly appear in 47.5% against 32.7%. Neither of those is something a model can generate from a topic alone, because both require a fact about your life or your reader that only you have.

So the useful mental model is: you supply substance, the model supplies structure. Every prompt below is built that way. If you want the deeper version of this argument, how to make AI posts sound human takes it apart properly.

What should you prompt for when you have nothing to say?

The blank page is the real bottleneck, not the drafting. These extract raw material rather than generating it.

1. The week audit.“I am going to describe my last working week in rough notes. Ask me eight questions that would help you find the two most post-worthy moments in it. Do not write anything yet.”
2. The repeated explanation.“Here is something I explain to clients almost every week: [paste]. What are three angles on this that a peer would find non-obvious rather than basic?”
3. The changed mind.“I used to believe [X] and now I believe [Y]. Interview me about what changed, one question at a time, until you have enough for a post.”
4. The disagreement.“Here is a widely held view in my industry: [paste]. Help me articulate precisely where I disagree and what evidence I have. Push back if my reasoning is thin.”
5. The mistake.“Ask me about a decision I got wrong in the last year. Then help me work out what the transferable lesson is, without making it saccharine.”
6. The number.“Here is a number from my work: [paste]. What is surprising about it to someone outside my company, and what context would they need?”
7. The archive.“Here are my last 10 posts: [paste]. What themes am I circling that I have not addressed directly?”

How do you prompt ChatGPT to draft the post itself?

8. The core draft.“Write a LinkedIn post from these notes: [paste]. Rules: open in first person with the specific moment, not a general observation. Do not resolve the story before the fourth line. Address the reader as ‘you’ at least once. No em dashes. No hashtags. Ban these words: leverage, delve, seamless, robust, unleash, elevate, testament, landscape, journey, game-changer. Under 200 words.”
9. The story shape.“Turn these notes into a post with this shape: the moment, what I expected, what actually happened, what it cost, what I would tell someone facing the same thing.”
10. The contrarian.“Draft a post arguing [position]. Steelman the opposing view in one line before dismantling it. Do not strawman.”
11. The teardown.“Here is something I read: [paste]. Draft a post that summarizes the useful part and adds my own view, so it stands alone without anyone clicking a link.”
12. The list post.“Turn these notes into a list post where each item is a specific thing I actually did, not a generic tip. Cut any item I could have copied from another post.”
13. The short one.“Compress this into under 60 words without losing the specific detail. Keep the strongest concrete noun in the first line.”

How do you prompt for a hook that earns the second line?

The visible fragment before the “see more” cut is roughly 210 characters on desktop and less on mobile, and it decides whether anything else gets read. Generate the opener separately from the body and generate many.

14. Ten openers.“Here is my post: [paste]. Write 10 alternative opening lines. Each must be under 180 characters and must not resolve the story. Number them.”
15. Find the buried one.“Read this draft and find the most arresting sentence anywhere in it. Move it to the top and rewrite the transition.”
16. The fold test.“Cut my post at 210 characters and show me only that fragment. Would a stranger tap to read more? Answer honestly and say why not.”
17. Styles.“Rewrite my opener five ways: as a first-person moment, as a direct address to the reader, as a flat declaration, as a genuine question, and as a number with a story behind it.”
18. The de-hype pass.“My opener sounds like a LinkedIn cliche. Rewrite it so it sounds like something I would actually say out loud to a colleague.”

Once you have candidates, run them through the hook analyzer, which checks them against the same patterns this article is built on and shows you exactly where the fold lands.

How do you strip the AI tells out of a draft?

This is where most of the quality gain is, and it is the step people skip. Editing prompts are more reliable than generation prompts because the model is checking against a rule rather than inventing.

19. The tell hunt.“Find every phrase in this draft that reads as AI-written. Look for hedging, rule-of-three padding, the ‘not just X, it is Y’ construction, throat-clearing openers, and abstract nouns doing work a concrete one should do. List them, then fix them.”
20. The dash sweep.“Remove every em dash and en dash. Replace with periods, commas, colons, or parentheses. Use plain hyphens for ranges.”
21. Specificity audit.“Mark every sentence that could have been written by someone who was not there. Rewrite each one to include a detail only I would know, or cut it.”
22. Read aloud.“Which sentences would sound strange said out loud? Rewrite those and leave the rest alone.”
23. Cut 30%.“Cut this by 30% without losing a single specific fact. Padding goes first.”
24. The claim check.“List every factual claim in this draft. For each, tell me whether it came from my notes or whether you introduced it.” This one catches invented statistics, which is the most damaging failure mode.

How do you get ChatGPT to write in your voice?

25. Build the voice guide.“Here are 10 things I have written: [paste]. Describe my voice as a set of rules another writer could follow. Include sentence length, how I open, what I never say, and my level of formality.” Save the output and paste it into every future prompt.
26. Voice check.“Here is my voice guide: [paste]. Does this draft match it? Point to the specific lines that do not.”
27. The de-corporate pass.“Rewrite this the way I would explain it to one smart colleague over coffee. Keep every fact.”
28. Comment drafting.“Here is a post I want to comment on: [paste]. Draft three replies that add something the author did not say. No compliments, no agreement without addition.”

What does a good prompt look like next to a bad one?

The gap between prompt levels is larger than any wording trick, so it is worth seeing it rather than being told about it. Same underlying event in all three cases.

Level 1: the topic prompt

Write a LinkedIn post about the importance of saying no to clients.

What comes back is structurally fine and completely inert: an opening about how saying no is hard, three reasons boundaries matter, a closing question. Nothing in it could only have been written by you, which means nothing in it is a reason to follow you. This is the prompt most people use and the reason most people conclude AI writing does not work.

Level 2: the specific prompt

Write a LinkedIn post about turning down a $40k contract last month because the client wanted us to skip the discovery phase.

Much better, because a specific fact is now in the room. The failure mode shifts: the model will pad the story with invented detail. It will guess at your reasoning, add a conversation that never happened, and put a tidy lesson on the end. You now have to edit out fabrication rather than blandness, which is more dangerous because it is less obvious.

Level 3: the material prompt

Here are my rough notes: [turned down $40k contract; client wanted to skip discovery; we did that once in 2024 and the project ran 3 months over; I said no on a Tuesday call and felt sick about it for two days; they came back six weeks later and accepted the full scope]. Write a LinkedIn post from these notes only. Do not add any fact I have not given you. Open in first person with the moment. Do not resolve it before the fourth line. No em dashes. Under 200 words.

This produces something usable in one pass, because every specific in the output came from you. The constraint that does the most work is “do not add any fact I have not given you”. It converts the model from an author into an editor, which is the role it is actually good at.

What no prompt can fix

Three limits are structural, and knowing them saves a lot of wasted iteration.

That third one is the quiet reason prompt-based workflows tend to decay. The first week is good because you are pasting careful context. By week six you are pasting a topic, and the output is back to the average of the internet. The honest comparison between prompting, tools, and hiring a person is in ghostwriter vs AI tool.

What are the parts of a prompt that actually works?

Collecting prompts is less useful than understanding why the good ones work, because then you can write your own for situations no list covers. Every prompt above that produces something usable contains four parts, and the weak ones are missing the third.

1. Role and reader.Who is writing and who is reading. “You are helping a fractional CFO write to founders of seed-stage companies” narrows the output more than any instruction about tone.
2. Constraints.Length, banned vocabulary, no em dashes, no hashtags, one idea. Constraints do more work than adjectives. “Under 150 words” beats “keep it punchy” every time.
3. Material. The actual thing that happened, with specifics. This is the part people skip and the reason most prompt lists disappoint. A model given a topic produces the average of everything written on that topic, which is by definition unremarkable. A model given your material produces something only you could have posted.
4. Output shape. Ask for ten options, or a draft plus three alternative openers, rather than one finished post. Judging is easier than writing, and you are a better judge than the model.

The asymmetry in that last point is the practical core of prompting well. You are not good at generating twenty openers, and the model is. The model is not good at knowing which one sounds like you, and you are. Structure every prompt so each side does the part it is better at.

A useful diagnostic: if you could send your prompt to a competent stranger and they could not write the post from it, the model cannot either. Missing material is almost always the problem, and no amount of prompt engineering substitutes for it. That is also why the question of what to write about is upstream of prompting entirely.

Which model should you actually use for this?

The prompts above work with any current frontier model. The differences that matter for LinkedIn writing are narrower than the benchmark arguments suggest, and they are mostly about default style rather than capability.

The last row is the one people get wrong. Chat memory feels like it is learning your voice, and what it usually does is average your voice with the model's. A voice guide you own, pasted deliberately, beats accumulated memory you cannot inspect.

Model choice is a small lever compared with input quality. A weaker model given your actual notes beats a stronger one given a bare topic, which is the whole argument of the worked example above.

Does AI-written content get penalised on LinkedIn?

No published LinkedIn policy penalises AI-assisted writing, and LinkedIn ships its own writing assistance, which would make such a rule awkward. What actually costs you reach is content that reads as generic, because the ranking half of the feed predicts human response, and readers scroll past writing that sounds like everyone else's.

That distinction matters for how you use these prompts. The risk is not detection. The risk is producing competent, unobjectionable posts that nobody has a reason to stop for. In our analysis of 12,988 posts, the sharpest divide between the top decile and the bottom half was comments (a median of 72 against 5), and comments are what generic writing never earns. That is the bar a prompt has to clear. We cover the detection question and the evidence behind it in AI content detection on LinkedIn, and the specific edits that remove the machine register in making AI writing sound human.

Do you have to disclose that a post was AI-assisted?

On LinkedIn, no. There is no platform rule requiring you to label an AI-assisted post.

Under EU law the answer is more specific than most coverage suggests. The AI Act's transparency obligations in Article 50 begin applying from 2 August 2026. The provision people worry about covers AI-generated text published to inform the public on matters of public interest, and it carries an exemption where the content has undergone human review and a person or organisation holds editorial responsibility for it. A personal LinkedIn post about your own work is not usually public-interest reporting, and a post you reviewed and published under your own name sits inside the editorial-responsibility exemption in any case.

This is general information rather than legal advice, and the answer changes if you are publishing news-like content at scale or operating in a regulated sector. The practical posture that survives every version of these rules is unchanged: read what goes out under your name, and be willing to stand behind it.

How do you build a prompting workflow that lasts?

1. Keep a running notes file. One line per interesting moment, captured the day it happens. This is the input everything else depends on.
2. Build your voice guide once using prompt 25, and reuse it.
3. Draft from notes with prompt 8, never from a bare topic.
4. Generate ten openers with prompt 14 and pick one yourself. Do not let the model choose.
5. Run the tell hunt and the claim check. The claim check is not optional if the post contains any numbers.
6. Shape it in the composer, where you can see the real fold. See line breaks.