You paste a prospect's profile into a chatbot, ask for a connection note, and get something polite, balanced and completely forgettable. It sounds like a brochure. You send it anyway because you have forty more to do. Three weeks later your reply rate has quietly halved and you cannot work out why.
The issue with AI LinkedIn messages is not that they are badly written. It is that they are written in the average voice of the internet — and the average voice of the internet is exactly what people have learned to ignore in their inbox.
You can fix this. Not by writing everything by hand again, but by giving the model enough of you to imitate, and by editing in a way that takes thirty seconds rather than three minutes.
Why AI drafts drift towards the same tone
A language model produces the most probable next word. Probable, across billions of documents, means corporate-neutral. That is why unguided drafts all converge on the same handful of moves:
- An opening compliment about the person's "impressive journey" or "great work at X".
- A pivot sentence beginning with "I noticed" or "I came across your profile".
- A three-part benefit claim with a tricolon rhythm.
- A closing question that offers "a quick 15-minute chat" with no stake in it.
- Hedges everywhere: "might", "could potentially", "I'd love to".
None of these are wrong. All of them are familiar. Your prospect has received the same shape forty times this quarter, and pattern-matching is faster than reading. The message dies before the second line.
The second drift is subtler. AI smooths out your specifics. You tell it you help operations directors cut invoice disputes; it gives you back "streamline your financial workflows". Every edit it makes moves from concrete to abstract, because abstract is more probable. That is the thing you have to actively reverse.
Build a voice file before you write a single prompt
A voice file is a plain text document — a page, no more — that you paste at the top of every prompt. It is the single highest-return thing you can do this week.
Put five things in it.
1. Ten sentences you actually wrote. Pull them from messages that got replies, from LinkedIn comments, from emails to clients you like. Not your website copy: website copy is already laundered. You want the way you write when you are explaining something to one person.
2. Your banned words. Mine, as an example of the format: delve, leverage, unlock, journey, passionate, synergy, elevate, game-changer, "I hope this message finds you well", "reaching out", "circle back". Write fifteen. Be ruthless. The ban list does more work than any instruction about "tone".
3. Your sentence rules. Something like: maximum 18 words per sentence; no sentence with three commas; contractions always; never two adjectives in a row. Models obey mechanical rules far better than adjectives like "friendly but professional".
4. What you actually sell, in your own clumsy words. Not the positioning statement. The way you describe it to a friend at dinner. "I help finance teams stop losing three days a month to chasing approvals." Clumsy and specific beats polished and vague.
5. Three things you never claim. No guarantees, no "proven system", no numbers you cannot evidence. This keeps the model from fabricating credibility on your behalf, which it will cheerfully do.
Save it. Reuse it. Update it when a message lands particularly well.
A prompt structure that produces drafts worth editing
Most people prompt in one block: "Write a LinkedIn message to this person." That gets you the average. Instead, separate the three inputs the model needs.
``` [VOICE FILE — paste in full]
FACTS ABOUT THIS PERSON:
- Role, company, company size
- One thing from their last 30 days (post, job change, hire, launch)
- Why I think they have the problem I solve
TASK: Write a connection note under 280 characters. One observation, one reason I'm writing, no question. Do not compliment them. Do not mention my services. Give me 3 versions with different first lines. ```
Three things matter here.
The facts are yours, not the model's. AI cannot see the profile unless you give it to it, and it will invent plausible details if you leave gaps. Spend ninety seconds gathering real ones — there is a repeatable way to research a prospect fast that gives you exactly the three facts this prompt needs.
The negative instructions carry more weight than the positive ones. "Do not compliment" removes a whole failure mode. "No question" stops the fake-urgency close.
Asking for three versions is not laziness. First drafts cluster around the mean; version three is usually where something unexpected appears. You will often take the opening of one and the body of another.
The thirty-second edit that makes AI LinkedIn messages yours
Never send a first draft. But do not rewrite it either — that defeats the purpose. Run three passes, in this order, and stop.
Pass one: kill the first sentence. Nine times in ten the AI's opening line is warm-up. Delete it and start on sentence two. This single move does more for your reply rate than any other edit. If what is left does not earn attention, you need a different angle entirely — the thinking behind openers that earn the second line applies to messages as much as to posts.
Pass two: swap one abstraction for one specific. Find the vaguest noun phrase and replace it with something only you could have written. "Improve your hiring process" becomes "stop losing candidates between the second and third interview". One swap is enough. Two starts to sound like a pitch.
Pass three: break a rule. Start a sentence with "And". Use a fragment. Leave a sentence slightly too blunt. Human writing has texture; smoothness is the tell. One deliberate roughness per message is the cheapest authenticity signal there is.
That is it. Thirty seconds. If you find yourself on a fourth pass, your prompt was wrong, not your draft — go fix the voice file.
Volume without turning into wallpaper
Here is where most people lose the thread. They build a good voice, get good replies at twenty messages a week, then scale to a hundred and the voice evaporates — because at a hundred, personalisation collapses into a merge field.
Two honest constraints.
First, the personalised variable has to be something you actually looked at. A company name pulled from a field is not personalisation; everyone knows the shape of {{company}}. One sentence that proves you opened their profile is worth more than four paragraphs of generated flattery. If you cannot write that sentence for a prospect, they do not belong on the list.
Second, follow-ups degrade faster than first messages. AI will happily produce a second and third message that simply restate the first in new words, and that is what gets you muted. Each follow-up needs a new reason to exist — a different angle, a resource, a clean exit. The structure in a three-message sequence for a no reply gives the model something to vary against instead of paraphrasing itself.
One more thing on scale. If you are considering a tool that sends on your behalf, be clear about what you are agreeing to. LinkedIn's User Agreement does not allow third-party software or browser extensions that automate activity on the platform. BIZLINKED is one such tool: its Chrome extension runs in your own browser and LinkedIn session and automatically sends invitations and messages, with first messages and follow-ups drafted by AI, inside limits it sets itself — at most 100 invitations a week, 150 once more than 40% are accepted, business hours only, random pauses. Those limits reduce exposure. They do not make it permitted. Accounts can be restricted at any volume, and anyone using automation is accepting that risk knowingly. Decide that with your eyes open rather than because a landing page told you it was fine.
What to do Monday morning
Open a blank document. Paste in ten sentences you actually wrote that got a reply. Add fifteen banned words. Write your offer the way you would say it out loud to a friend. That is your voice file, done in twenty minutes.
Then take the next five prospects on your list. Gather three real facts each. Run the prompt structure above, apply the three-pass edit, and send.
At the end of the week, read the five messages back without looking at who wrote them. If you cannot tell they are yours, the voice file is too thin — add more sentences, not more instructions. And if your replies have been thinning out for months regardless of the tool, the problem may be upstream of the writing: the reasons outreach templates stopped working are rarely about word choice.
AI will not give you a voice. It will only amplify the one you hand it. Hand it a good one.
