Particles and WavesPAW
AI × Engineering

The code AI writes, humans can't review

The faster you can ship, the less anyone understands what shipped. A note on the productivity paradox.

Scroll
🤖
Symptom 01

PRs got bigger

You ask the AI to fix one bug; it returns a 500-line diff. It works. Tests pass. But nobody on the reviewing side can really follow what's happening anymore.

Symptom 02

More 'LGTM' without reading

Approving without reading becomes the norm. The code gets written, but no one on the team holds the whole picture in their head.

Symptom 03

Future-you can't read it either

When you come back to fix something AI wrote, you can't read your own diff. You end up asking the AI to explain it. You've given up ownership of understanding.

73%

Output vs. comprehension

Speed of writing goes up dramatically with AI assist. Self-reported codebase comprehension drops, six months in, for 73% of small teams I've talked to.

Informal, small sample / illustrative
Code is read ten times more often than it's written. We're paying for the writer's efficiency with the reader's cognitive load.
A line I keep coming back to
🪞
A mirror — staring back at you
Workflow

How you use AI decides the outcome

Anti-pattern

The 'write it all' workflow

One big prompt, one big diff, approve when tests pass. The model does the thinking; your understanding doesn't grow.

Recommended

The 'write it together' workflow

Have the model produce small chunks (tens of lines). Read each one, understand it, then ask for the next. You lose speed, but you keep ownership of comprehension.

1/3

A workable balance

Cutting speed by a third in exchange for keeping ownership of understanding. That's the trade I've landed on, for now.

Thanks for reading

More essays on this kind of thing

Particles and Waves is where I publish what I've learned wrestling with AI tools as an indie developer — on the blog, and in short Kindle books.