Is ‘Vibe Coding’ Killing Code Quality? Bram Cohen’s Take on AI‑Driven Development

Bram Cohen, the founder of the BitTorrent protocol, recently published a sharp critique of Anthropic’s Claude team, accusing them of over‑relying on a practice he calls “Vibe Coding,” which he says has caused a dramatic decline in software code quality.

“Vibe Coding” refers to a newly emerging AI‑driven development approach where developers depend entirely on AI‑generated code, avoid writing or even reviewing the underlying logic, and advance projects through vague conversations with the AI.

Cohen deems this approach “absurd beyond reason.” He stresses that even when developers do not write code themselves, they must still establish a solid foundation—project planning documents, a backlog, skill modules, and rule systems—otherwise AI efficiency plummets and core tasks may fail.

Following the Claude code leak, the developer community discovered extensive duplication and redundancy in the codebase. Cohen questioned why Claude’s developers never inspected the code themselves, noting that within the “Vibe Coding” culture, looking at the underlying code is treated as a form of cheating.

He highlighted obvious problems such as definitions that were simultaneously labeled as both “Agent” and “Tool,” which he described as redundant and in need of cleanup.

According to Cohen, the true value of AI programming is not to turn developers into “hands‑off managers” but to help teams efficiently eliminate technical debt. He explains that traditional software projects can accumulate years‑worth of debt, which AI‑assisted refactoring can resolve in weeks instead of a year of manual effort.

Cohen also shared his own workflow: he first engages the AI in thorough discussions to clarify project edge cases, potential risks, and core requirements, then directs the AI to perform specific refactoring tasks, ensuring each step has a clear objective.

He concludes with a provocative insight: using AI does not obligate developers to accept low‑quality software; poor code results from the choices developers make, and they must take responsibility for those choices.