I Had Three AIs Write My Voice. They Didn't Agree.

The disagreement turned out to be the most useful part.

Three AIs wrote me a voice guide. Claude, Gemini, and ChatGPT each took a pass. I figured I’d skim them, keep the best lines, and be done by lunch.

They didn’t agree. Not on tone, not on what I write about, not even on who I am.

That sounds like a problem. It was the most useful thing that happened all week.

Here’s why they disagreed. Two of the three had built their guide around an older version of me, heavy on monetization and growth tactics. All real parts of my background. None of it where I’d actually landed. What I’m building now runs on a simpler rule: no public pitch. I’m learning in public, and the writing isn’t a sales funnel. The AIs weren’t wrong. They were working off old game film.

So the job changed. I wasn’t averaging three documents. I was deciding which version of me was current and reconciling everything to it. One draft had the sharpest voice. The other two had the better scaffolding, the templates and the vocabulary. I kept the sharp part, kept the useful scaffolding, and cut everything that was selling a business I’m not running in public.

Then came the rounds. And this is the part I didn’t expect: every pass made the guide less about adjectives and more about control. The edits that mattered weren’t “sound warmer.” They were the boring, load-bearing ones:

• Rule priority. An order for what wins when instructions collide, because they will. Accuracy and privacy beat sounding like me, every time.

• A missing-receipts rule. “Proof over hype” is a great line right up until the AI invents the proof. So if there are no receipts, write it as a hypothesis or a question. Never fake a case study.

• “Context is not boilerplate.” Every draft wanted to open with “as a non-engineer spending 20 hours a week...” Once, that’s context. Every time, it’s a tic.

• A prompt an agent can actually run, split into the part that never changes and the part that’s today’s assignment.

By the end it was about 580 lines. I QA’d it by running it back through all three AIs until they stopped finding things to fix. So I now own a 580-line document about how to sound less like a 580-line document. Progress.

If you want to find out where your own positioning has quietly drifted, here’s the cheapest test I know: have three different AIs describe you, and watch where they disagree. The gaps are the map.

For the record: I built the guide with AI, and used AI to help write this post about building it. The voice is mine. The typing was a team effort.