On vibecoding

I have 2-4 Claude terminals open at any given time. Different parts of the same project, sometimes a second project running alongside. I describe what I want, Claude writes the code, I read it, think “LGTM,” and move on.

That’s the honest version.

It feels like crime sometimes. Like cheating. Like I’m the dog in the burning building drinking coffee, convinced everything is fine because things are shipping.

The thing I keep coming back to

When something breaks, I don’t dig in. I describe the problem to Claude and let it fix it. I ask why — I do that — but the answer always feels like something I’m approving rather than something I understand. LGTM without really knowing.

What scares me isn’t that I’m using AI. It’s that I don’t know if I’m learning anything. I can tell you what my code does. I can’t always tell you how.

Vibe coding vs vibe engineering

There’s a version of this that I think is actually good. Where you understand the architecture deeply, you know the patterns, and you use AI to move faster within that understanding. You’re steering. The AI is just executing.

That’s vibe engineering.

What I’m sometimes doing feels more like vibe coding — where Claude is both the architect and the builder and I’m mostly just reviewing PRs I don’t fully understand.

I’m not sure which one I’m doing on any given day. That’s probably the thing worth sitting with.

Is this just how it is now

I don’t think I’m alone in this. I just think most people aren’t saying it out loud. There’s a lot of “shipped X in Y hours” energy and not a lot of “I have no idea what half of it does” energy.

Maybe that’s fine. Maybe the bar is shifting and what matters is whether the thing works and helps people. Maybe deep architectural knowledge is becoming less important and product sense is becoming more important.

Or maybe I’m just rationalizing.

I genuinely don’t know. But I think it’s worth being honest about not knowing.