I found out something uncomfortable a couple weeks ago: an AI can sound completely fine and still not sound like me. Not wrong, not broken, just... generic. It took me longer than I'd like to admit to notice, because "fine" is a hard thing to get mad at.
The Problem With "Fine"
Here's what happened. I was backfilling the blog on my photography business site, rexjones.photo, which had gone quiet for way too long. I had an agent read a couple of my published posts and match the voice from there.
Ten posts came back. They were fine. Correct facts, reasonable structure, nothing embarrassing. And they read like literally any small business blog. Helpful, friendly, a little bland. If you'd shown them to me without my name on them, I wouldn't have known they were supposed to be mine.
Why Two Samples Aren't Enough
The thing is, I hadn't actually given the agent much to work with. Two sample posts tells you the topic and the general shape of a sentence. It doesn't tell you the actual rules I write by, because those rules live in my head, not on the page in any obvious way.
So the agent did the only reasonable thing it could do: it wrote competently. It just optimized for the wrong target. It was matching "a" voice, some plausible business-blog voice in that general neighborhood, instead of matching my voice specifically.
And that distinction matters more than it sounds like it should, because those two things produce genuinely different writing even though both versions are technically correct.
The Document I Already Had
Here's the part that actually embarrasses me a little: I'd already solved this problem. Months ago, I sat down and wrote out, in real detail, how I actually write. Not vibes, not "be punchy" — actual rules.
Specific rhetorical patterns I use, and how often is too often for each one. A rule that metaphors have to be things you could touch, not corporate abstractions — nothing that sounds written, everything sounds like it could come out of my mouth mid-conversation. A sentence rhythm I use on purpose, where a sentence builds and builds and then lands hard on something short.
I had a list of words I don't use, ever, because they sound like they came out of a press release. Even a rule that before I go after someone else's mistake, I have to admit I've made that exact mistake myself first. I even wrote a second document whose whole job is to read my own writing back like a hostile editor and flag anything that sounds AI-generated.
That document existed the entire time. I just didn't hand it to the agent doing the writing. I gave it two sample posts and assumed that was enough context to reverse-engineer a voice I'd spent real effort making specific. It isn't enough, and honestly it was never going to be enough — two samples show you where I've been, not the rules that got me there.
What Actually Changed
Once I actually pulled that document up and used it, the difference wasn't subtle. Same facts, same topics, same structure even — but suddenly it sounded like something I'd actually say instead of something adjacent to what I'd say.
I ended up going back and rewriting all ten of those first posts from scratch against the real rules. Every post I've written since started with that document open, not as an afterthought but as the actual brief.
The Real Lesson
The lesson isn't "AI can't sound like you," because it clearly can, mine does now. The lesson is that "sounding like a specific person" and "sounding professionally competent" are different jobs, and if you only give an AI enough to do the second one, that's the job it's going to do. It's not going to guess its way to the first job on its own.
If you've ever gone to the trouble of writing your own voice down — actually written it down, not just told yourself you'd know it if you saw it — the fix for generic output usually isn't a cleverer prompt. It's just remembering the document exists and actually using it.
