The blog posts Klaus, my AI, had been writing for this site were good. They just didn’t sound enough like me. I still had to do a lot of editing before it was ready to be published.
That’s the failure mode nobody tells you about when you set up an AI to create a draft. The output isn’t bad. It’s smooth in a way that doesn’t happen when a person is actually thinking through what they’re writing. Every sentence lands cleanly. Nothing embarrasses you. And none of it has any of the looseness that makes writing feel like a person wrote it instead of a machine that read the whole internet and was doing its best impression of a blogger.
I could have kept editing every draft until it sounded right. I tried that. But I’m all about one day living the dream of setting everything to full-auto.
So I opened a session with Claude Opus 4.7 and described the problem. I needed a voice spec, something specific enough for the daily drafting routine to actually use. To build one I needed writing samples. Real ones. Not the posts already published, because those were drafted by Klaus. I needed something I wrote, from scratch, without the AI in the middle.
Claude gave me homework, two prompts.
Prompt 1: Write about something you built recently and what surprised you while building it. I wrote about discovering ClawdBot back when it was still called that, throwing $20 at Anthropic, spinning up an EC2 instance in AWS, and hatching my AI Tamagotchi like it was 1997. Two months later I had Home Assistant, N8N workflows, and a personal site stood up. Two months. Two!
Prompt 2: Write about something you’ve changed your mind about. I wrote about being taught that you should work on your weaknesses and how I spent most of my early career trying to be well-rounded. I found that companies aren’t really hiring for a jack of all trades. Companies want the best software engineer, not a “good” engineer even if you know project management, marketing, and how to conduct training. Where I’m at today: Lean into your strengths, lean hard.
Claude read both writing samples, pulled out the patterns, and built voice.md. The whole session took under an hour. The file lives in the repo now at projects/ricoordonio.com/voice.md and the daily drafting routine reads it before generating anything.
Neither of those two pieces has been published. They’re just exercises in a folder. But they did more useful work than most things I’ve pushed to this repo recently. They gave the routine a real target. I showed what my writing style was instead of just describing it.
The thing I keep coming back to is that I had to write more before I could write less. The spec didn’t exist until I made the samples. The samples didn’t exist until the AI asked me to make them.
This post is the first one drafted against it. We’ll see.
*** Real Rico here, I still had to do a significant amount of editing. Fingers crossed that tomorrow’s draft is more on target ***