← All posts

How to write LinkedIn DMs in your voice (when you've been letting AI do it)

April 28, 2026 · 7 min read

If you've been outsourcing LinkedIn replies to ChatGPT for the last year, the writing muscle that gets your DMs read has probably atrophied a bit. This post is the rough framework I use to find my voice again — concrete enough to use, loose enough to bend.

Quick prerequisite: if you've never had a "voice" in writing, that's a different post. This one assumes you used to write your own messages, then started letting GPT do it, and noticed your reply rate slipped.

The five tells

These are the patterns recipients now register, often unconsciously, as "AI wrote this":

  1. Opener clichés. "I hope this message finds you well." "Trust this finds you well." Even "Hope you're having a great week." Real friends don't open messages with these. Drop them entirely.
  2. The 3-paragraph shape. Para 1: flattery. Para 2: pitch. Para 3: ask for time. This shape is so over-used that recipients pattern-match it before reading words.
  3. Corporate verbs. "Leverage." "Synergize." "Circle back." "Touch base." The 2010s sales-deck vocabulary. People hear them and disengage automatically.
  4. Over-precise CTA. "Would 15 minutes on Tuesday at 2:00pm PT work?" The over-specificity reads as a calendar bot. Real humans propose ranges, then coordinate.
  5. Emoji density. 🚀 ✨ 💪 in the first message from a stranger. AI tools default to friendly emoji as politeness lubricant. People have caught on.

The framework I use

Five principles. Not rules. Bend any of them when the situation calls for it.

1. Start with one specific observation

Generic flattery ("loved your post!") is read as bot. A specific observation ("the bit about chunk overlap in your RAG post made me rethink how I was indexing emails") is read as human, because no AI bothered.

The unit of authenticity is specificity. One concrete reference is worth ten generic compliments. If you can't make a specific observation, you probably haven't done enough homework to be sending the DM.

2. Use contractions. Most of them.

Compare these:

I am writing to ask whether you would be open to a brief conversation.

I'd love a quick chat if you're up for it.

Same content, different humans. The first one was written by a robot trying to be professional. The second was written by an actual person. Contractions cost nothing and signal humanity.

3. End with one ask, lightly held

Pick one. "Want 15 min sometime next week? Totally fine to pass." Not "Could we discuss a partnership opportunity, schedule a meeting, and align on next steps?"

The "lightly held" part matters. Adding "no worries if not" or "totally fine to pass" shifts the dynamic from request to invitation. Replies go up because you've removed the social cost of saying no.

4. Match the length of their last message

If they sent you 3 sentences, write back 3 sentences. If they sent you a long detailed question, you can write 5–6 sentences, but probably not 12. The length match is a cheap signal that you're paying attention and not burning their time.

AI defaults to 4–5 sentences regardless. Your reply will stand out by being shorter when the context calls for short, longer when it calls for longer.

5. Keep one verbal quirk

Have one habit that's distinctly yours. Mine is em-dashes — I use them more than most people do, and they signal "this is the same person who wrote the other 5 DMs in this thread." Yours could be lowercase first letters, or always saying "appreciate" instead of "thank you," or starting messages with "Quick one."

Recipients build a mental fingerprint of your style across messages. If the AI version of you is registerless, the inconsistency itself is a tell.

Before / after

Same situation, two messages.

Before (AI-flavored)

Hi Sarah,

I hope this message finds you well. I came across your impressive work in the AI evaluation space and was particularly inspired by your recent insights on multi-agent systems. I'm reaching out to explore a potential synergy — I'm building a tool in an adjacent area and would love to leverage your expertise.

Would you be open to a brief 15-minute conversation next week? I'm flexible to align with your schedule.

Looking forward to your response.

Best,
Alex

After (mine)

Hey Sarah, appreciate you reaching back. Quick one — I'm building a small tool that helps recruiters draft DMs in their own voice instead of generic AI copy. The bit in your last post about evaluation latency is exactly the problem I'm running into. Would love 15 min to hear how you'd think about it. No pitch, genuinely just learning. Free anytime next week?

Same intent. Different reply rate.

The shortcut

If you've been writing the AI-flavored version for a year, just doing the inverse takes deliberate effort. The shortcut I built into Reply Coach is to feed the model 5–10 of your own past replies as style references — not just instructions. The model doesn't need to be told "use contractions, drop clichés, add em-dashes." It just imitates the patterns in your samples.

That's the whole product, basically. You paste 5 past replies once, the rest of your outreach matches their cadence forever.

Either way — whether you write your DMs by hand or use a voice-trained tool — the north star is the same. Your reader should be able to tell, in two sentences, that a person wrote it. The whole reply game in 2026 is downstream of that.

Reply Coach learns from 5 of your past replies and writes new ones in your voice. 5 free per day, no card required.