AI Workflow

Every compulsive behavior I’d ever heard warned about had one thing in common: it was obviously a waste. Junk food, endless scrolling, the worse stuff a person can lose an evening to — the harm announces itself. You know what you’re doing while you do it. The internal argument is short because the verdict is never in question. You’re not building anything. You’re feeding a loop. The recognition that it’s destructive is what eventually lets you stop, or never start. ...

Back in January I argued that AI parallelism turns clean architecture into a forcing function — coupling now has an immediate cost in blocked parallel work. After three more posts living inside this workflow, I want to come back to that argument and say what I think I underestimated. The economic shift is bigger than parallelism. AI has quietly collapsed a whole stack of tradeoffs that software engineering treated as fundamental, and the implications run further than “you can run more agents.” ...

The previous post was about when to go up a level. This one is about the operational layer underneath that decision: which model to use for which kind of work, and the harder discipline that makes the whole thing actually work — being willing to throw away expensive output. The two topics are connected. If you’re not willing to discard, your model choices don’t matter, because you’ll end up patching whatever the cheapest model produced regardless of how wrong its foundation was. ...

In the first post I argued that responsibility in AI-assisted development isn’t line-level review — it’s strategic oversight at the right altitude. This post is about what that means in practice. Specifically: how to recognize when the right move is to stop pushing at the current level and step up a layer. This is the single most useful pattern I’ve found, and it shows up at every scale. The Pattern Whenever something isn’t working — a bug AI keeps failing to fix, a feature it keeps implementing wrongly, a plan that keeps generating awkward code — the wrong response is to push harder at the same level. The right response is almost always to go up one. ...

A few months ago I wrote a post arguing that AI parallelism would force developers to take clean architecture seriously, because coupling now has an immediate cost in blocked parallel work. I still think that’s true. But I want to start this series with a confession: the elaborate workflow I built on top of that insight is mostly not what I actually do anymore. This isn’t a retraction. The economic argument holds. What I got wrong was the workflow that I thought followed from it. ...