I'm starting from zero. And I have no clear direction - apart from submitting to some generic directories.
Since I'm solo, I would love to just prioritize what's at least worked well for others.
I can get AI to write good code. That's not the problem.
The problem is everyone else thinks they can too.
Most of my week now is reviewing PRs from juniors. It compiles. Tests mostly pass. Looks fine at a glance. Then I actually read it.
Wrong abstractions. Doesn't match our architecture. Files thrown wherever the model felt like that day. And when I ask why something's built the way it is, there's no answer. Because they don't really understand it.
So review isn't "is this correct" anymore.
It's me reverse-engineering whatever the model spat out, figuring out why it's subtly wrong for our system, then explaining fundamentals the author never had to learn in the first place.
Half the time it's faster to just rewrite it myself.
And the stuff I miss? That's the stuff making the codebase fragile six months from now.
This isn't an "AI bad" post. The tool's great. I use it every day.
The issue is it lets people skip the part where you actually learn how to structure code and think about systems. Which is the exact thing review depends on.
So I tried to fix it.
I started encoding our conventions into a rule-based "how-we-ship" setup. Forced adherence through a /ship skill. Basically a compliance + elegance checker that runs before a PR can even go up. Push the standards left so they hit the guardrails before they hit me.
It helps. It catches the dumb structural stuff.
But it's not close to perfect. It still waves through the subtle design problems. The ones that need actual judgment. Which are, of course, the exact ones I care about.
Optimist in me says this gets better. Models get smarter, rule systems get sharper, the slop floor rises.
Pessimist in me says we're raising a generation of devs who never build the judgment to catch any of this themselves.
Anyway. For the experienced folks here:
- You seeing this too? Review load going up while junior output "looks" better than ever?
- Anyone actually gotten automated checks to catch design problems and not just style? Or does it always plateau at the surface stuff?
- Is this temporary, or is this just what senior eng is now?