The Guy Who Built Claude Code Says Don't Use It for Important Code

Boris Cherny, the Anthropic engineer who built Claude Code, went on record this week to explain when AI coding works well — throwaway code, prototypes, stuff not in the critical path — and when you still want a human being thinking carefully about every single line.

That second category is called "software."

I don't want to be uncharitable here. The quote is honest and the caveat is real and it's genuinely refreshing when someone selling a tool cops to its limits. But there's something clarifying about watching the creator of the thing draw the boundary at "maintainable code" — as if maintainability were a niche preference, a vintage affectation, like caring about code review or writing tests or whatever other practices we used to do before we figured out you could just yell at a chatbot until something compiled.

The dream was always that AI would handle the boring parts. The boring part, it turns out, is the part where you think carefully about what you're building and why. The fun part — the prototyping, the velocity, the throwaway experiments — AI is quite good at that.

We have successfully automated the wrong half of the job.