Do not index
Why do your AI-written LinkedIn posts still sound like everyone else's no matter how many times you rework the prompt? Because the prompt was never the problem. The input is. A model given nothing but a topic and a tone instruction will produce the statistical average of every post ever written on that topic, and the statistical average is exactly what your feed is already full of. Jodie Cook made this case in Forbes recently, and her line cuts through a year of prompt-engineering noise. "You probably need better input, not a more advanced prompt," she writes. I run content at Hivemind and use AI daily across client accounts, and I can tell you from inside the workflow that she is right.
This article is for founders writing their own posts, ghostwriters charging $5k to $30k per month, and content leads at 3 person agencies trying to keep a dozen client voices distinct while using the same tools for all of them.
Skip this if you are looking for a magic prompt to copy. There isn't one, and if you are still shopping for prompts instead of building an input pipeline, this article will not change your results. It will only explain why they are flat.
Cook's suggested fix is feeding the model your profile, your best old posts, voice-note transcripts, and your actual rants, then editing until you would say it out loud. That maps almost exactly to how professional ghostwriting already works. The reason ghostwritten content at a good agency does not sound like AI is not that the writers avoid AI. It is that the process starts with extraction, getting the raw opinions, stories, and speech patterns out of the client's head before anyone drafts a word. No prompt can substitute for that step because the material does not exist anywhere a model can reach it.
How to make AI LinkedIn posts sound human
I think about inputs as what I call the Extraction Spine, four layers of raw material that have to exist before drafting starts. The first layer is identity, your positioning, who you serve, and what you are known for. The second is proof, your 10 best-performing posts, because they carry your cadence and your audience's demonstrated taste. The third is speech, transcripts of you actually talking, from voice notes, client calls, or podcast appearances, since your spoken syntax is the fingerprint text-only inputs miss. The fourth is heat, the live opinions and rants you have this week, because conviction cannot be retrofitted into a draft. Feed a model all four layers and it drafts like a competent assistant who has studied you. Feed it none and it drafts like the internet.
At Hivemind the spine gets built through interviews. A 30 minute conversation with a client produces more usable material than 3 hours of prompting, because the questions surface stories the client did not know were interesting. Founders writing their own content can run the same play solo. Record a voice note on your commute answering one question about your week, transcribe it, and hand the transcript to the model as the source. The draft that comes back will contain your phrases, your examples, and your actual position, which is material a prompt could never have invented. This is also why the strongest founder content comes from people who treat their day-to-day work as the raw material, the same logic behind positioning yourself as a practitioner first rather than a thought leader.
The editing pass still matters, and this is where most people quit early. Cook's standard is reading the draft aloud and rewriting anything you would not say. Mine is slightly stricter. If a sentence could appear in a competitor's post without anyone noticing, it goes. What survives that filter is short, and that is the point. A 200 word post that is 100% yours beats an 800 word post that is 60% average.
The part of Cook's piece that deserves to be printed and taped to every content team's wall is her other line. "I don't care if AI wrote your content. I care if it's good," she writes in Forbes. The AI-versus-human debate is a distraction from the only question audiences actually ask, which is whether the post said something worth their 30 seconds. Readers do not run detectors. They run pattern recognition, and the pattern they punish is emptiness, not automation.
The strategic implication is bigger than any single post. Every transcript, rant, and interview you bank becomes a proprietary dataset that compounds, and it is the one asset in your content operation nobody can copy. Two writers with identical tools will produce increasingly different results based entirely on the quality of the raw material they have collected about themselves or their clients. The people who spent this year hunting better prompts own nothing. The people who spent it building their Extraction Spine own a library that makes every future draft faster, sharper, and harder to imitate. That gap widens every week, and it is quietly becoming the difference between commodity content services and the ones that keep their retainers.
