The AI writing tool I built so I would stop sounding like a press release

I tried a handful of AI writing tools to help draft LinkedIn posts. Every single one handed back something that sounded like a press release.

Technically correct. Completely generic. Not me.

“Write in my voice” is not an instruction

Most tools I tried asked for one line of style guidance, something like “be warm but direct,” and then generated a full post from that single sentence. The output was fluent. It was also nobody in particular.

Here is the thing I learned building my own version: a style note is not enough because “write in my voice” is not an instruction, it is a specification. A model has no idea what warm-but-direct actually sounds like without examples of what it does and does not sound like. That gap, between a vague style note and an actual specification, is where most AI content ends up sounding like nobody.

What I built instead

So I built Post Writer. Instead of a one-line prompt, it works from a written voice profile, tone, vocabulary, sentence rhythm, the phrases I actually use and the ones I never would. It suggests topics grounded in live web search instead of stale training data, drafts the post against that profile, and shows me the sources behind every suggestion so I can check a claim before I post it.

My actual workflow looks like this:

1. Subject: me
2. Research: AI
3. Draft: AI
4. Finished post: me

The tool picks the topic candidates and does the first pass. I still decide what is worth writing about, and I still decide what actually goes out.

Humans stay in the loop, even on the post about the tool

It is only ever a starting point. I edit every draft before it gets anywhere near the publish button, this article included. Some of my posts are 100 percent written by me with no AI draft in the loop at all. Humans must remain in the loop (HITL) wherever a wrong call is expensive, and getting a fact or a tone wrong in public, under my own name, counts.

It is the same standard I hold client builds to. A model that drafts something is useful. A model that publishes on its own, with nobody checking the output against reality first, is a liability wearing a demo. If you are weighing where an AI tool should stop and a person should pick up the pen, that is exactly the kind of boundary I help set. You can find the details on the contact section of this site.