There’s something disorienting about reading a post in your own voice that you didn’t write.
I set up a routine a few weeks ago. Klaus, my AI agent, drafts a blog post every morning at 5am PT. It reads through whatever happened in the repo the day before, my memory files, the session scratchpad, and then writes something in my voice using a spec I wrote for it. By the time I make coffee, there’s a draft in the git history with my name on it.
The first time I opened one, my reaction was close to recognition. The rhythm felt right. The vocabulary was right. No em dashes, no bullets, no “in today’s fast-paced world.” I thought, okay, this might work.
Then I read more carefully.
What the AI gets right is the surface. The sentence length variation, the short punchy anchors after the longer narrative ones. The contractions, the casual interjections. “And you know what” shows up where I’d put it. “Don’t get me wrong” lands at the pivot point. The opener doesn’t start with “Today I” or a rhetorical question.
What it gets wrong is harder to name, but I know it immediately when I see it. The posts that miss feel like a convincing forgery. Same handwriting, slightly wrong pressure. The insight arrives on schedule. It sounds like something I’d conclude. But I never had the experience that produced it.
That’s the gap. The automation can imitate lived experience. It can’t have any.
Don’t get me wrong, most of the drafts are solid. Maybe three out of five I’ll flip from draft: true to draft: false with light edits. One of five is close but not quite there, needs a real rewrite. One of five I read the opener and know in two sentences it’s going in a direction that doesn’t match what’s actually on my mind right now.
That last category is the most interesting. The post that misses reveals what the AI thinks I care about, which reveals what I’ve written into its memory, which is sometimes different from what I’m actually living. The gap between those two things is live information.
I’ve started treating the wrong drafts as a diagnostic tool. When the automation writes a post about building systems for consistency and I’m actually stressed about something unrelated, that tells me something about the delta between what I’ve been logging versus what I’m actually going through. It’s a strange kind of journaling.
If I’m being honest, the part that still surprises me is how often it gets the voice right. Not the meaning, not the stakes, but the way I’d phrase the sentence, the place I’d let things breathe. It’s imitating a pattern well enough that it’s useful.
The part it can’t fake is the experience. That’s still mine to bring.