You paste your talking points into ChatGPT. You get a post back. It starts with "In today's fast-paced world, professionals are discovering new ways to..." You delete it. You start over. You try again with a slightly different prompt. The second draft opens with "I'm excited to share that..." You close the tab.
This is the loop most people are stuck in. The problem is not that AI cannot write. The problem is that without specific constraints, AI writes like the average of every LinkedIn post it has ever seen. That average is very, very generic.
This article gives you a repeatable system: a voice input you build once, a prompting structure that uses it, and an editing workflow that turns a competent draft into something that reads like you wrote it. No more deleting and starting over.
Why AI-written LinkedIn posts sound robotic
This is a prompting problem, not a technology problem.
Large language models predict the next most likely token given everything before it. That is the whole mechanism. When you give a model no context about your voice, your audience, or your opinion, it fills the gaps with its training defaults.
Its training defaults for LinkedIn skew toward formal, hedged, corporate language. Motivational endings. Five-bullet frameworks. Rhetorical questions that beg for engagement. The model is not being lazy. It is being accurate. This is what LinkedIn posts statistically look like.
The fix is not a better model. The fix is making the model's defaults impossible to reach by giving it stronger constraints than its training data.
What the model is actually doing
The model is not trying to sound like you. It is trying to sound like the average of every LinkedIn post it trained on. Your job is to make that average impossible to reach.
Build your voice input before you write a single prompt
A style guide is the prerequisite. Without it, every prompt starts from zero.
A style guide is a short document, 200 to 400 words, that tells the AI how you write, not what you write about. You build it once. You paste it into every prompt. It is the difference between a model that sounds like LinkedIn and a model that sounds like you.
The guide covers four inputs. Each one does different work. Skip one and the model will fill that gap with its defaults.
Sentence length
How long your typical sentences run and whether you vary them or keep them short.
Vocabulary register
The actual words you use, not the level of professionalism you aim for.
Opinion style
How directly you state a point of view and how much you hedge.
Story triggers
The types of examples and moments you draw from when making a point.
Voice extraction prompt
Claude / GPT-4Here are 5 LinkedIn posts I've written in the past. Analyze them and write a 200-word style guide that describes: - My average sentence length and rhythm - Words or phrases I repeat - How I open posts (do I use questions, statements, or stories?) - My opinion style (direct, hedged, conversational?) - What I never do (bullet lists, rhetorical questions, motivational endings) Write the style guide in second person, as instructions for a writing assistant. Start with "When writing as [Name], you should..." [PASTE YOUR 5 POSTS HERE]
The prompting system that actually works
Three layers. Each one does different work. Skip one and the output degrades.
Most people write one-line prompts: "Write a LinkedIn post about [topic]." This gives the model nothing to work with except its defaults. A production-grade prompt has three layers: role and style guide, raw material, and output constraints. The model uses all three simultaneously.
Here is how each layer functions and what happens when you leave it out.
The three-layer prompting structure
Role + style guide
Who you are, how you write
Raw material
The idea, story, or opinion to write about
Output constraints
Format, length, tone rules
Draft
AI output, two versions
Edit pass
You make it real
Set the role
Paste your style guide and tell the model it is your writing assistant, not a content creator. "You are a writing assistant for [Name]. Do not generate ideas. Do not add context I didn't give you. Write only from the material below." This single instruction prevents the model from inventing examples you didn't provide.
Drop in raw material
Give the AI the actual substance: a rough paragraph, a voice memo transcript, a bullet list of observations from a client call. The rougher the better. AI polishes. It should not invent. If you give it nothing, it will fabricate something plausible and generic.
Add hard output constraints
Specify what the post must not do. "No bullet lists. No rhetorical questions. No motivational ending. No sentences over 18 words. Do not use the word 'excited'." Negative constraints are more reliable than positive ones. Telling the model what to avoid is more precise than telling it what to aim for.
Request two variants
Ask for two different opening lines. This forces the model to explore the material rather than defaulting to the first structure it finds. It also gives you a real choice. Editing a draft you partly like is faster than rewriting one you don't.
Mark what to keep
Before you edit, highlight one or two lines from the draft that sound most like you. These become your anchor points. The edit pass is about pulling the rest of the post toward those lines, not rewriting everything from scratch.
Full production prompt
Claude / GPT-4You are a writing assistant for [Name]. Your job is to turn raw material into a LinkedIn post. Do not add ideas, context, or examples that aren't in the material below. Do not make the post sound inspirational or motivational. Style guide: [PASTE YOUR STYLE GUIDE HERE] Raw material: [PASTE YOUR ROUGH NOTES, TRANSCRIPT, OR BULLET POINTS HERE] Output rules: - No bullet lists - No rhetorical questions - No sentences over 18 words - Do not start with "I" or "In today's" - Do not end with a call to comment or engage - No em dashes - Write in plain, direct language Write two versions of this post. Each version should have a different opening line. Label them Version A and Version B.
What to do with the draft
Three edit moves that turn a competent AI draft into something that reads as human.
AI prose has a texture problem. It is too smooth. Every sentence connects cleanly to the next. Every transition is logical. Every paragraph lands its point. Real human writing has friction: a sentence that trails off, a word repeated for emphasis, a thought that pivots mid-paragraph.
Three specific moves fix this. They take less than ten minutes. They change the post more than any amount of reprompting will.
Add friction. Break one sentence in two. Remove a transition word. Let a thought stand alone without explanation. The small roughness signals a real person wrote this.
Cut the scaffolding. AI loves to announce what it is about to say. "Here's what I've learned:" or "The key takeaway is this:" Delete these. Start with the thing itself. The setup adds words and removes punch.
Insert one specific detail. Add one fact, number, name, or moment that only you could know. "The client was on their third agency in two years." "This was a Tuesday in February." Specificity is the fastest signal of a human writer. AI cannot generate your specific details. Only you can.
The most common edit mistake
The most common mistake in the edit pass is adding words, not removing them. AI drafts are usually 20% too long. Cut before you add. Every sentence you remove makes the sentences that remain hit harder.
Building a style guide that scales across a team
How to handle voice differentiation when writing for multiple people.
Individual style guides work for one person. Teams, agencies, and ghostwriters managing multiple personal brands need a layered structure. The goal is the same: give the model enough constraints that it cannot default to generic. The structure just has more layers.
The three-layer architecture below separates what stays constant from what changes per person. Company voice is the base. Role voice sits above it. Individual voice sits at the top. Each layer adds specificity.
Team voice architecture
Company voice
Role voice
Individual voice
How to write a voice guide for an executive you're ghostwriting for
You need 30 minutes and a recording. Ask these four questions and let the executive talk. Do not interrupt. Do not summarize. Record everything.
- "Tell me about a decision you made recently that you'd do differently." This gets you a real story with a real opinion. It is the raw material for three posts minimum.
- "What's something most people in your industry believe that you think is wrong?" This surfaces their actual point of view, not the one they perform in meetings.
- "How would you explain what you do to someone who has never heard of your company?" This reveals vocabulary, analogies, and the level of jargon they actually use.
- "What's a word or phrase you hate seeing in business writing?" This is gold. Whatever they say goes directly into the output constraints section of every prompt you write for them.
Record the answers. Paste the transcript into the voice extraction prompt from Step 1. The style guide it produces will be more accurate than anything you could write by hand.
Prompt templates for the four most common LinkedIn post formats
Copy, fill in the brackets, and run. Each template is pre-loaded with constraints that block AI defaults.
Four formats cover the majority of high-performing LinkedIn content: the opinion post, the story post, the observation post, and the list post done without the usual five-bullet structure. Each template below is pre-loaded with the constraints that prevent the most common AI failure modes for that format.
Fill in the bracketed sections with your actual material. The rougher your input, the more the constraints matter. Do not clean up your raw material before pasting it in.
Template 1: The opinion post
Claude / GPT-4You are a writing assistant. Write a LinkedIn post that states a direct, specific opinion about [TOPIC]. Do not hedge the opinion. Do not balance it with "on the other hand." State the opinion in the first sentence. Then give one specific reason why you hold it. Then give one example or observation that supports it. No bullet lists. No rhetorical questions. Under 150 words. Style guide: [PASTE HERE] Opinion: [STATE YOUR ACTUAL OPINION IN ONE SENTENCE] Supporting reason: [ONE REASON] Example: [ONE SPECIFIC THING THAT HAPPENED OR THAT YOU OBSERVED]
Template 2: The story post
Claude / GPT-4You are a writing assistant. Write a LinkedIn post that tells a short story. Start the story in the middle of the action, not at the beginning. Do not open with context-setting. Do not explain what the reader is about to learn. End the post with one sentence that states what you took from the experience. No bullet lists. No motivational language. Under 200 words. Style guide: [PASTE HERE] What happened: [DESCRIBE THE SITUATION IN ROUGH NOTES OR BULLET POINTS] What you did or said: [THE SPECIFIC ACTION OR MOMENT] What you took from it: [ONE SENTENCE, YOUR ACTUAL CONCLUSION]
Template 3: The observation post
Claude / GPT-4You are a writing assistant. Write a LinkedIn post based on something I noticed or read recently. The post should state the observation in the first two sentences. Then explain why it matters or what it signals. Do not add general advice. Do not end with a call to action. Write as if you are thinking out loud, not teaching. Under 120 words. Style guide: [PASTE HERE] What I noticed: [THE SPECIFIC THING YOU SAW, READ, OR HEARD] Why it matters: [WHAT IT SIGNALS OR WHAT IT MADE YOU THINK] Context: [WHERE YOU SAW IT, WHEN, ANY RELEVANT DETAILS]
Template 4: The list post done right
Claude / GPT-4You are a writing assistant. Write a LinkedIn post structured as a short list. Each item in the list should be a complete, specific sentence, not a label followed by an explanation. The list should have between 3 and 5 items. No item should start with the same word as another. No em dashes. No item should be a generic principle. Each item should be specific enough that it could only apply to the context I give you. Open with one sentence that frames the list. Do not close with a summary or call to action. Style guide: [PASTE HERE] Topic: [WHAT THE LIST IS ABOUT] Items: [YOUR RAW POINTS, AS ROUGH AS YOU LIKE]
The pattern across all four templates
Every template does the same three things: it tells the model what structure to follow, it tells the model what to leave out, and it requires you to supply the actual substance. The model cannot write a good opinion post if you don't give it your actual opinion. The templates force that input.
