Hacker News debates where AI-assisted code stops holding together

A Hacker News post pulled 519 points and 510 comments in 20 hours, arguing AI removes the friction that used to force engineers to think. The piece, "Coding expertise is going to collapse from AI reliance", turned its comment section into an argument about who loses the skill.

The top comment named the contradiction. Employers say writing code by hand means doing it wrong, while the result arrives faster than anyone can review it line by line.

Several commenters pointed to a specific breaking point. Codebases built from AI output nobody supervised start falling apart somewhere between 10,000 and 50,000 lines. Below that range, one person can still hold the design in their head. Above it, nobody can, human or model.

The thread also settled on what models still can't do. State a negative constraint and make it hold. Telling an agent what to build works. Telling it what never to do, and having that rule hold once the agent is several files deep, is a different problem. I write CLAUDE.md files with exactly that kind of rule in my own projects, and I've watched a model follow the rule in the file it's editing and ignore it in the next file it touches.

That gap needs a check that runs against the diff. No prompt file rule reliably survives a model that's already moved past the file where it was written.