§ 01 · What we builtFive posts a week. Zero writing days.
Every Sunday afternoon I take a one-hour walk and record voice notes — usually 10–14 of them, two minutes each, about whatever happened that week in the business. By Monday morning, five of them are live on LinkedIn as fully-formed posts. I never sit down at a keyboard to "write a post."
The whole thing runs on a Zapier flow that watches a Notion table. I tag a voice note "ready"; an hour later it's drafted in my voice; I approve it with one click; Buffer schedules it.
§ 02 · Problem being solvedLinkedIn is a tax I have to pay for inbound.
I run a small B2B SaaS. About 30% of my pipeline comes from LinkedIn. I've measured it. I don't want to post — I want the inbound that comes from posting. So the problem was never "how do I get better at writing LinkedIn posts." It was "how do I never have to think about LinkedIn between Monday and Friday."
§ 03 · What was the hardest partMaking the drafts sound like me, not like a bot.
The default ChatGPT voice is the kiss of death on LinkedIn. Every post sounds like a corporate training module. The fix: I fed it 40 of my old posts as a style reference and built a very specific anti-style prompt — no emojis at line starts, no rhetorical questions, no "Here's what I learned," no em-dashes (yes, ironic), no three-word punchy openers.
It took about three weeks of tweaking before the drafts felt like me. After that, the editing time per post dropped from 15 minutes to 90 seconds.
§ 04 · What went wrongThe approve tag stopped propagating.
§ 05 · What you learnedThe hard part is noticing what's worth writing about.
One: the AI can write the post. It can't notice the moment. Voice notes during the walk are where the actual value comes from — the system is just a transcription/polish layer.
Two: style references beat style prompts. Showing 40 old posts works infinitely better than describing "my voice" in adjectives.
Three: always have a human-in-the-loop checkpoint. The day I removed the approval step, two posts went out that I would have killed. Don't.