I pieced together a workflow from half-remembered blog posts (there are too many ways to coerce the machine), and to my surprise it mostly worked. Once I got going, the weird part is that I felt more like an editor than a developer. Sure, I was steering the conversation (half paying attention and focused on other important things like letting the dog out) but mostly I was just pressing enter, not furiously typing out rebuttals like I've seen in many posts.
The strange part is that I tell the thinking machine what I want, hope for the best, and most of the time, I get what I want. Because I don't get what I want every time, the whole thing feels like the model is quietly saying "I have altered the deal, pray I don’t alter it any further". And when I'm surprised with with what I didn't ask for, I laugh and say to myself "this deal is getting worse all the time" and keep going.
Despite this, I'm enjoying Claude Code and will likely continue the subscription beyond my pilot period.