Bram Cohen Says the Cult of Vibe Coding Is Insane. 569 HN Points and 463 Comments Agree.

BitTorrent creator Bram Cohen published "The Cult Of Vibe Coding Is Insane" and it hit 569 points with 463 comments on Hacker News. The most-discussed vibe coding essay to date.

Cohen's trigger was the Claude Code source leak. When the 512,000 lines of TypeScript leaked via a missing .npmignore, people laughed at the code quality. Cohen's argument: the Claude team went "completely overboard with dogfooding," refusing to even spend a few minutes looking under the hood. The result was code that worked but was messy in ways a few minutes of human review would have caught.

His thesis is simple. Pure vibe coding, where you deliberately refuse to look at the implementation, is "dogfooding run amok." Cohen describes his own approach: start by asking the AI to audit code, discuss problems until actionable solutions emerge, explain what should change, then let the machine write. He reads the code. He has opinions about the code. He just doesn't type it.

I agree with about 80% of this. The remaining 20% is where it gets interesting.

Cohen is right that human review is cheap and high-value. Five minutes of reading a diff catches structural problems that an AI will cheerfully propagate across an entire codebase. My design background makes me care about this instinctively. You don't ship a design you haven't looked at. Why would you ship code you haven't read?

But I think Cohen undersells what pure vibe coding is good for. Prototyping. Throwaway scripts. One-shot tools you'll use once and delete. When the cost of a bug is "I wasted 10 minutes," human review adds overhead without proportional value. The skill is knowing which mode to be in.

The connection to Story #1 is direct. Claude Code's quality regression means vibe coding works less well this week than it did two months ago. The people who were reading their diffs noticed the regression faster than the people who weren't. Cohen's point lands harder when the AI is actively getting worse.

What builders should do: read your diffs. Seriously. Not every line, but the structural decisions. Does the file organization make sense? Are the abstractions appropriate? Is the error handling real or theatrical? Five minutes of human taste applied to AI output is the highest-leverage activity in software engineering right now.