Bram Cohen doesn't mince words. The BitTorrent creator looked at Anthropic's leaked Claude Code source and found redundant agents and tools that any human reviewer would have caught. His diagnosis? The team fell into "vibe coding" dogma, refusing to peek under the hood.
Vibe coding, as Cohen describes it, means letting AI write everything while treating code review as "cheating." The real issue: refusing to engage with what the AI produces. Cohen points out the Claude code is written in plain English. Anyone could read it and notice duplication. Instead, the team apparently believed pure prompt-based development makes manual oversight obsolete, fundamentally changing the development paradigm.
Cohen's alternative is what he calls "Ask mode." You talk through problems with the AI first, clarify edge cases, correct its wrong assumptions. Then, once you've done that back-and-forth, the AI can execute cleanly. He's been working this way for months, starting conversations with "this function makes my eyes bleed" and iterating until there's a concrete plan. The AI does the implementing. The human provides direction.
His bottom line is blunt: "Bad software is a decision you make." AI can help clean up technical debt faster than ever, but only if you point it at the mess. Refusing to look at your own codebase isn't a methodology. It's abdication.