How to write LinkedIn posts that don't sound like AI
Change the order: judgment → sources → one draft you can reject
Why tip lists and prompt packs fail
Most people searching this query get the same advice: soften the tone, add a personal story, use this prompt pack, chase a detection score. That advice treats the symptom. The posts that read as “AI” usually fail for a simpler reason: nobody owned the claim before the model wrote the sentences. The draft is fluent. It still sounds like no one in particular. “Sounds like AI” is almost never about whether a model touched the text. It's whether a busy reader believes you wrote it — and whether you'd defend it if someone pushed back.
Direct answer
Write LinkedIn posts that don't sound like AI by changing the order of operations. Start with a judgment you would defend in a room — a claim, a refusal, a tradeoff you actually live with. Then pull sources only as evidence for that claim. Then produce one draft you are willing to reject out loud. Run a short keep/kill: keep the line you'd stand behind; cut the fluent empty line. If you can't say “that's me” or “that's not me,” you don't have a writing system yet — you have a reading habit with extra steps. Better prompts won't fix a missing judgment.
The wrong starting place
Prompt-first LinkedIn writing looks efficient. You open a chat. You type: write a post about X. You get something clean. Sometimes clever. Then you try to sand it into something you'd put your name on. That order is backwards. If you haven't decided what you think, the model invents a writer. It invents the average one: safe hooks, generic lessons, tidy takeaways, no scar tissue. Tip lists make the same mistake. They optimize for polish. Polish without judgment is how you get polished emptiness — the thing people now casually call slop.
What actually works: judgment → sources → draft
1. Judgment first — Before any tool, answer out loud: What am I willing to defend in a room? Something you shipped. Something you refuse. Something your market keeps getting wrong. A tradeoff you live with. If you can't say the claim without hedging, don't prompt yet. You don't need a better model. You need a sharper point. Voice is not a polish pass at the end. It's a constraint at the start: what you refuse to say, what you always cut, what you'd put your name on. 2. Sources second — as evidence, not as progress. Pull only what supports the judgment you already named. If a source doesn't change the claim or the evidence under it, leave it out. More inputs will not fix a missing judgment. They will delay the draft that would have exposed it. 3. One rejectable draft — Generate once toward the claim, not toward “a good LinkedIn post.” The useful moment is meaner than people want: a draft you are willing to reject out loud. If you can't say “that's not me” or “that's close,” you don't have a writing system yet. You have setup. The honest job of a writing product is to get you to a draft you can reject. Everything before that is optional scaffolding.
A 5-minute keep/kill
1. Read the draft out loud once. 2. Keep one sentence you'd defend in a room. 3. Cut one line that is fluent but empty. 4. Stop when you can name why those two decisions — not when it “sounds done.” Fluency is a trap. AI drafts are good at sounding finished. Smooth sentences. Clean transitions. Nothing to trip over. That's why they survive review when they shouldn't. The filter isn't “does this read well.” It's “would I say this out loud, and would I defend the claim.” Run the keep/kill on meaning first. Fluency can wait.
What to skip
Prompt packs as strategy — they rearrange words around an undecided point of view. “Write in my voice” prompts with no prior judgment — the model still has to invent the writer. Do not treat detectors as the goal — a passing score is not the same as owning the claim. Fix the writing system, not the score. Auto-posting — if the system ships for you, it also skips the only judgment that made the draft yours.
Try the workflow once
If you want to practice this without turning it into another research project: pick one claim you'd defend, attach only the sources that matter, generate one draft, and keep/kill it out loud. No signup pitch required to understand the point. The point is the order: judgment → sources → rejectable draft.