You post for your work. Two or three times a week on LinkedIn, maybe elsewhere, because it brings in clients or candidates or readers, and because someone decided you should. You have also watched the feed fill up with posts that all have the same shape: a confession in the first line, a line break after every sentence, a lesson, a question at the end. You know an assistant wrote most of them. You do not want to be the next one.
Social posts are the shortest format on this site and the one where AI tells are most concentrated. A 120-word post has nowhere to hide a hook formula or a stack of one-line paragraphs. The model has read millions of these posts and has learned their skeleton perfectly, which is exactly why its drafts are recognizable at a glance. The skeleton is the problem, and the skeleton is the first thing you have to take away from it.
The loop still applies, but the weights are unusual. Directing is mostly supplying real material and a platform-specific banned list. Proposing is fast and cheap, which is a trap: ten posts in one turn means ten copies of one post. Judging is mostly about the first line and the tells. Owning means two things here: everything the post says happened to you did happen, and the replies underneath it are written by you.
What changes for this format
The method is the same everywhere. These are the places where this format bends it.
The first line is the post
Every platform truncates, so the first line is what most readers see and all they use to decide. The model writes first lines it learned from the feed: I got rejected 47 times. Here is what I learned. Unpopular opinion. Nobody talks about this. I almost quit last year. These are hook formulas, and the readers you want have been trained by the feed to scroll past them. Your first line should be a specific sentence you would say to a colleague who asked what happened this week. It is usually a fact with a number or a name in it, and it is usually the line the model buried in the middle of its draft.
Platform tells: the broetry stack and its friends
You know the shape. One sentence per line. A blank line after each. Let that sink in. Here is the thing. A list of three with arrow bullets. A rhetorical question. Agree? The model produces this on request and often without one, because it is the dominant shape in the data. Your brief needs a banned moves list specific to the platform, beyond banned words: no single-sentence paragraphs, no line break after every sentence, no lesson paragraph, no question as the last line, no arrow bullets, no hashtag block. Ask for the post as two or three normal paragraphs. A paragraph has to hold a thought together. A stack of lines does not, which is why the stack is so easy to generate.
Batch variation without sameness
Ask for ten posts in one turn and you get ten posts with the same skeleton, varied by topic. That is the sameness readers detect across a month of your feed even when each post looks fine alone. The fix is to vary the job each post does rather than the wording. One answers a question a client asked. One admits a mistake and what you changed. One is a number and what it meant. One quotes a conversation. One disagrees with something you read. Draft them in separate turns, each with its own raw material, and never ask for variations of a post that already works.
The invented anecdote
This is the one that can cost you. Ask for a post about leadership and the model will write: when I was 24 I was fired from my first job, and it taught me. You were not. The model does not know your biography, so a story in the first person that you did not supply is fiction wearing your name and your headshot. On a professional network, that is a reputational matter, and people do check. The rule is absolute: every event in a post happened, and the details come from your source note. If the model offers a story you did not give it, cut the story and tighten the brief.
Replying in your own voice
Comments are where the humans are, and where readers find out whether anyone is home. An assistant-generated reply is usually longer than the comment it answers, thanks the commenter for their insight, and restates the post. People recognize it instantly. Write replies yourself, short, and specific to what the person said. If you draft a reply with the assistant, it should be shorter than the comment and should contain something only you could know. Mostly, do not draft them at all. The posts can be assisted. The conversation is you.
Disclosure norms
Platforms are adding labels for AI-generated media, employers are writing policies about posts under their name, and professional norms are still forming. The honest default is simple to state: if the ideas and the facts are yours and you edited every line, the post is yours, and a general note on your profile about how you work is a courtesy rather than a confession. If the model wrote the story, the opinion, or the numbers, the post is the model's whatever the label says. Check your employer's policy before posting under their name, and follow the platform's rules on labels where they apply.
The process
- 01You
Keep a running source note of real material
Social posts die for lack of material, never for lack of prompts. Keep one running note: the question a client asked on Tuesday, the number from the report, the mistake and the fix, the sentence from a conversation that stuck. Each entry is a post's worth of raw material. When you sit down to draft, you are choosing from this list, not asking the assistant what to post about. If the list is empty, the problem is upstream of the chat.
- 02You
Brief with a platform-specific banned list
Use the LinkedIn and Social mega prompt. The part that does the work is the banned moves list, and it should be specific to the platform you are posting on, with the formulas you see most often in your own feed. Add your voice notes. Add the rule that nothing in the post happened unless it is in the source note. Ask for one post per turn.
Send this Hold these as fixed for every post in this session. Banned moves: single-sentence paragraphs, a line break after every sentence, any first line that is a confession or a number of failures, unpopular opinion, here is the thing, let that sink in, a lesson paragraph, a closing question, arrow bullets, a hashtag block, emoji. Every event must come from my notes; invent nothing. Write in two or three ordinary paragraphs, under [WORD LIMIT] words. One post per turn, then wait.
- 03AI proposes
Draft one post per turn, each with its own job
Pick one entry from the source note, name the job (answer the question, admit the mistake, explain the number), paste the material, and ask for one post. Not variations. The draft's job is to give you a shape to react to and a middle you can mine for the first line. Expect to keep the middle and replace the ends.
- 04You
Rewrite the first line yourself
Find the most specific sentence in the draft, which is usually in the middle, and move it to the top. Or write a new one: the sentence you would say to a colleague. Then delete whatever the model put first. This single move does more for a post than any other edit, and it is almost never something the assistant will do on its own, because its training says the hook goes first and the fact goes later.
- 05AI proposes
Bounded edits for tells and length
With the first line fixed, use Bounded Edit for what remains: rejoin stacked lines into paragraphs, cut the lesson, remove the closing question, bring it under the limit. Name the move rather than the vibe, and ask for OLD and NEW so you can see exactly what changed and keep the sentences that were already yours.
Send this Change only the structure of this post, not the facts or my sentences. Rejoin the single-line paragraphs into two paragraphs. Delete the final paragraph that states the lesson and the closing question. Keep my first line exactly. Bring the whole post under [WORD LIMIT] words by cutting, not compressing. Show OLD and NEW, then wait.
- 06AI proposes
Critique for sameness in a fresh chat
Once a week, paste the week's posts into a fresh chat and run Critique Your Draft with one question: which lines could appear in anyone's post, and which two posts share a skeleton. Do not let it rewrite. Use the answer to adjust the banned list and to notice which job you keep defaulting to. This is the only critique you need for this format, and it is the one that catches the slow drift toward the feed's average.
- 07You
Post, then reply by hand
Before posting, read it once more as a skeptical peer and confirm every event in it happened. Post it. Then handle the comments yourself: short, specific, no thanks for sharing. If your employer or platform has a disclosure rule, follow it. If not, a line on your profile about how you work is enough. The replies are where people decide whether the account is a person. Keep that part.
Toolkit
The prompts that fit this format, and when to reach for each.
- Mega promptLinkedIn and Social mega promptThe session set-up. Its banned moves list and one-post-per-turn rule are the whole defense against the feed's skeleton.
- Stage promptBounded EditFor rejoining stacked lines, cutting the lesson, and hitting the word limit without the model rewriting your first line.
- Stage promptCritique Your DraftWeekly, in a fresh chat, on the batch. Ask which lines could be anyone's and which posts share a shape.
Before and after
A paragraph the way an assistant tends to return it, and the same paragraph after a writer took it back.
I made a mistake last week. A big one. I sent a proposal to a client without double-checking the numbers. The result? We lost the deal. But here is the thing: Mistakes are not failures. They are lessons in disguise. Here is what I learned: → Always double-check your work → Own your mistakes → Move forward with humility Growth happens outside your comfort zone. Agree?
We lost a proposal last week because I sent it with last year's rates in the pricing table. The client noticed before I did, which is the part I keep thinking about. I have a checklist for proposals. It has nine items, and the numbers match the current rate card was not one of them, because I had never gotten that wrong before. It is item one now. If you keep a checklist for anything, it is probably built from the mistakes you have already made. The next one is going to come from somewhere the list does not cover.
The first line is now a specific sentence with the actual mistake in it, instead of a confession formula followed by a two-word paragraph. The stack of single lines became two ordinary paragraphs, which forced the sentences to connect to each other and exposed that the lessons-in-disguise paragraph had nothing in it. And the lesson is particular: checklists are built from past mistakes and miss the next one, a thought that came from the writer, replaced growth happens outside your comfort zone, and the closing Agree? is gone. Every detail, including the nine items, came from the writer's source note.
Pitfalls
The first line is a confession, a number of failures, or unpopular opinion.
Move the most specific sentence in the draft to the top, or write the line you would say to a colleague. Add the formula to the banned list so it stops coming back.
One sentence per line, blank line between each.
Ask for two or three ordinary paragraphs in the brief and rejoin by bounded edit. If a thought cannot survive being in a paragraph, it was not a thought.
Ten posts from one turn, all with the same skeleton.
One post per turn, each from its own source note entry with its own job. Never ask for variations of a post that works.
The post tells a story about you that did not happen.
Cut it entirely. Add every event must come from my notes to the brief, and check each post against the source note before posting. This is the one that costs reputations.
Replies under the post are long, grateful, and restate the post.
Write replies yourself. If you must draft one, it should be shorter than the comment and contain something only you could say.
A lesson paragraph and a closing question keep appearing.
Both are banned moves, so name them in the brief. End the post on the last fact or the last specific thought, and let people reply if they want to.
Questions
How do I make AI-assisted LinkedIn posts sound human?
Supply real material, ban the platform's formulas by name, ask for ordinary paragraphs, and rewrite the first line yourself. The sound of a human post comes from specific facts in a connected argument, which is what the stack of one-line paragraphs is designed to avoid. If a post still sounds generated after that, the usual cause is that the model invented the material. Replace it with something that happened.
Should I say a post was written with AI?
Follow your employer's policy and the platform's rules first. Beyond that, if the facts, the opinion, and the edits are yours, a general note on your profile about how you work is honest and sufficient. If you would be embarrassed for a reader to learn which lines the model wrote, that is a sign to rewrite those lines rather than to decide about a label.
Can I batch a month of posts in one session?
You can batch the drafting, but one post per turn, each with its own material and its own job, and ideally across more than one session. A single turn that produces twenty posts produces one post twenty times. Also expect the batch to go stale, because the best posts respond to something that happened this week.
Is it okay to use the assistant to reply to comments?
It is where readers most reliably notice, so the practical answer is no for most replies. A short reply you wrote in thirty seconds beats a polished one the model wrote in five. If you use the assistant at all, use it for a reply to a long or difficult comment, then cut it to half the length of the comment and make sure it contains something only you could know.