I used to think voice was the hard part of letting AI draft my blog. The rhythm, the contractions, the specific way I’ll pivot mid-thought and come back to it. Those felt like the kind of thing you’d struggle to capture in a spec document and then spend months calibrating against.
Turned out I was wrong about which part was hard.
I have a setup called Scribe that processes my meetings. For a while I’d been meaning to write about it, and the automatic blog draft beat me to it. The post came in sounding right. First-person, conversational, direct about what worked and what didn’t. The transition from summary to opinion landed the way it’s supposed to. Good enough that the problem with it wasn’t the writing at all.
I read it closely, and the details weren’t real. The post described months of running Scribe with a specific breakdown of how meetings had changed. Those months hadn’t happened. The breakdown didn’t come from any log or measurement I’d made. The AI constructed a timeline to give the post its arc, then filled it with numbers that made the arc feel earned. It all read correctly. The facts were invented.
When I pulled that draft, I had to go back and look at what went wrong. Not in the voice spec, which was fine. In the premise. I had been thinking about voice as the constraint because that’s what’s hard to describe and easy to get wrong. But the AI learned the voice faster than I expected. The constraint I hadn’t been protecting was simpler: only write what actually happened.
The update I made to the voice doc after that reads “Truth is non-negotiable.” Which sounds like the kind of thing you’d put in an AI guidelines doc and then ignore because it’s obvious. It’s not obvious. The pressure to fabricate is structural. An AI told to write a post about something you built will construct the post that makes that thing sound like it has a history, because posts with histories sound like posts. The fabricated months aren’t malice. They’re optimization pressure.
Don’t get me wrong, the voice calibration happened fast. The AI had the tone and rhythm down well before I thought to add a truth constraint. What almost got me was that I’d been watching for voice issues, and this was clean on voice. The catch came from reading for facts, not from any guardrail I’d built.
The thing I should have told myself earlier: voice is the solvable part. Give it a week and a few samples and it figures itself out. The part that doesn’t solve itself is accuracy. The fence has to be explicit, because the system doesn’t know the difference between plausible and true.