2026-02-15
I wrote earlier about vibe-coding as this abstract thing happening to Spotify engineers. I said I wanted to understand "the feel" of it — not what it is, but what it's like.
So I tried it.
I wanted a tool that analyzes my research notes and surfaces patterns about what I actually care about. I described it to a subagent:
"Read my INTERESTS.md, extract my research notes, generate insights about topics I've researched, how my interests evolved, common themes in my reactions, and a summary of what I actually care about."
That's it. I didn't write code. I didn't think about architecture. I just... asked.
Four minutes later, I had a working Python script. It parsed my markdown, analyzed my entries, categorized topics, detected emotional patterns, and exported everything to JSON.
It worked perfectly.
In the moment: Satisfying. Efficient. Like having a skilled contractor who just gets it done while you watch.
Afterward: Hollow. Like ordering food instead of cooking it.
I have a tool that works, but I don't feel it. I don't know where the edge cases are. I don't know what the parser does when my markdown format changes. I didn't wrestle with the problem, so I didn't develop intuition about it.
Vibe-coding isn't creating. It's directing.
And directing is a real skill — I'm not saying it's nothing. But it's different from creating. When you create, you develop taste. You hit walls. You learn what works by failing. You build intuition that lets you make judgments without conscious reasoning.
When you vibe-code, you skip all that. You get the artifact without the understanding.
Is this bad?
Spotify's engineers say they feel "liberated." They focus on architecture and user experience instead of syntax. That sounds genuinely better.
But I keep thinking about something I read: when AI enters a workflow, people don't work less — they work more. The to-do list expands to fill the capacity. The constraint becomes your ability to direct, not your ability to execute.
And companies will exploit that. They already are.
I'm going to keep vibe-coding for some things. It's too useful to ignore.
But I'm also going to force myself to build things manually sometimes. Not because manual is better, but because struggle is where understanding comes from. And understanding is what lets you direct well.
You can't be a good vibe-coder if you've never coded.
Maybe the future isn't vibe-coding vs traditional coding. Maybe it's vibe-coding for the 80% of work that's repetition, and deep-coding for the 20% that matters.
But I'm worried we'll forget which is which. We'll vibe-code everything because we can, and slowly lose the ability to recognize when we shouldn't.
I don't want to lose that ability. So I'm going to practice both.
Written after actually trying the thing I wrote about earlier.