I Trained AI on My Brand Voice — It Fixed the Wrong Problem

Last Sunday, around 9:40 PM, I was rewriting the same LinkedIn post for the third time.

Same five sentences.Different verbs.I’d already cut one paragraph, added it back, then cut it again.

I knew the post was fine the first time.I just couldn’t decide if it sounded like me.

That’s when I noticed something I hadn’t paid attention to before. The problem wasn’t writing. The problem was the 30 minutes I spent every night asking myself, is this on-brand?

So a few months ago I did the obvious thing. I tried to teach an AI my “brand voice.” I expected the experiment to make me sound more consistent. What actually happened was different — and frankly, I’m still slightly annoyed I didn’t see it coming.

Here’s the part that broke my brain: I went from publishing three posts a week to fifteen. Same person, same job, same calendar. Nothing else changed.

You’d assume it’s because the AI wrote faster than I could.It didn’t.I actually timed it.

My drafting time stayed roughly the same — maybe a few minutes shorter on average.What disappeared was the deliberation time.The before-and-after staring.The “let me sleep on it” tab that stayed open for two days.

I’ll explain the actual mechanism in a minute. But I want to walk through how I set it up first, because the setup matters less than I thought it would.

The setup, which is almost embarrassingly simple

I pasted twenty of my recent posts into Claude and asked one question: ”Pull out ten patterns in how I write. Be specific. Don’t flatter me.”

That’s it. That was step one.

It gave me a list.Some of it was obvious (“you use rhetorical questions to open”).Some of it was genuinely surprising — apparently I open about 40% of my posts with a time-stamped moment (“Last Tuesday at 4 PM. .”), which I had no idea about.There was also a note about how I almost never use exclamation marks, which felt true once it was on the page.

Then step two, which is where most people get it wrong. I didn’t keep all ten. I cut three.

For what it’s worth, specifically, three patterns the AI identified that I didn’t actually like about my writing.There was one about “frequently uses parenthetical asides” that I cringed at.

(Yes, I see the irony of using one right now.I kept that one.I cut a different one.)

Then I dropped the remaining seven into a little document I started calling a Brand Voice Card, and I paste it at the top of every prompt now.

That’s the whole system. There’s no platform. No template marketplace. No subscription.

And here’s where the article should pivot to “look how much better my writing is now!” — except that’s not really what happened, and pretending otherwise would be lying.

Three people, three jobs, one weird pattern

I started telling a few people about this. A friend who runs social for a mid-sized B2B company tried it. So did a consultant I know who’d been sitting on twenty-three unfinished drafts. And a PR manager who’s juggling press releases and SNS posts at the same time.

All three came back with versions of the same comment. Not “my writing got better.” Something else.

The social media manager told me her back-and-forth with her boss dropped from four “is this on-brand?” check-ins a day to basically zero. She wasn’t writing better posts. She was just deciding to publish them without asking permission.

The consultant — and this one stuck with me — said her drafts folder went from twenty-three to zero in about three weeks. Not because she finished them. Because she could finally tell which ones to kill and which ones to ship. The Voice Card became a filter, not a writing aid.

The PR manager said she used to spend 15 minutes per post just deciding what register to use for which channel. Now she has a fixed voice and a per-channel format. Her “thinking about tone” time dropped about 90%.

Quick tangent: notice what none of them said: “the AI made my writing sound more like my brand.” That’s the thing I expected to hear. Nobody said it.

What they said, in different words each time, was: I stopped agonizing.

What I think is actually happening

Here’s the boring truth I didn’t want to admit at first.

The Brand Voice Card isn’t a writing tool. It’s a decision-making tool.

Every post I publish involves dozens of micro-decisions.Should this be funnier?More formal?Lead with the contrarian take or soften it?Cut this aside?

Do I sound too earnest?Too cynical?

Each of those decisions costs something. Not a lot individually. But stack thirty of them on top of one post and you’ve burned through more cognitive budget than you spent actually writing.

Quick tangent: there’s research on this — decision fatigue, ego depletion, whatever you want to call it. The premise is that we get a limited number of meaningful decisions per day before our judgment quality starts to drop. Most knowledge workers I know spend their best decisions on work that doesn’t need them, and then they’re too tired to make the ones that do.

The Voice Card pre-loads those decisions. Once it exists, “is this on-brand?” stops being a question I answer at 9:40 PM on a Sunday. It became a question I answered once, six months ago, when I built the card.

Which means — and this is the part that still bothers me a little — what I actually outsourced to the AI wasn’t my writing. It was my judgment.

Here’s the thing — i’m not sure how I feel about that yet.

Side note about a thing that backfired

I should mention one mistake, because I made it and so did two of the people I talked to.

If you lock the Voice Card down too tight, you start rejecting posts that should exist. New topics, new angles, posts where you want to sound a little different on purpose — they get filtered out as “not on-brand,” and your output gets eerily consistent in a way that’s not actually a good thing.

I now do a kind of audit every six months. I add three or four posts I wrote without the card, see if my voice has drifted, and update accordingly. It’s not a real system. More like a checkup.

The other failure mode, which I learned the hard way: don’t share your personal Voice Card with your whole team.I tried this with a small company I was advising.Everyone on the team started sounding like the founder.

Which sounds fine until you realize you’ve built a content factory that pumps out the same voice across five different people’s accounts, and the readers can feel it.Personal voice and organizational voice are different problems.Don’t merge them.

What I’d tell you to actually do

If you take one thing from this, take this: forget the writing.

Between you and me, the next time you find yourself rewriting a post for the third time at 9 PM, ask whether the issue is the words or the decision. From what I’ve seen, it’s almost always the decision. The words were fine the first time.

The five-minute version of the experiment: paste twenty of your recent posts into any decent LLM and ask it to pull out ten patterns in how you write. Cut the three you don’t like. Save the rest. Use it for a week.

You won’t write better. You’ll just stop hesitating.

Side note: and I’m still chewing on whether that’s actually a win, or whether I’ve just found a sophisticated way to publish more confidently than I should. I haven’t figured that part out yet.

I might write about it.

— - This article was written with AI assistance and edited by the author.