Bram Cohen critiques 'vibe coding' and AI-assisted development practices

The problem with 'vibe coding'

Bram Cohen's article criticizes what he calls 'vibe coding'—a development approach where teams use AI coding assistants while deliberately avoiding looking at the underlying code. He argues this is 'dogfooding run amok,' where using your own product becomes a cult activity that goes beyond reasonable limits.

Claude source code example

The article cites Claude's source code leak as a case study. Cohen notes that when people examined the leaked code, they found significant duplication issues—specifically, 'a whole bunch of things that are both agents and tools.' He questions why developers didn't notice this themselves, attributing it to the 'vibe coding' mentality where 'looking under the hood is cheating.'

Practical AI collaboration approach

Cohen advocates for a more engaged approach to AI-assisted development:

• Start conversations with the AI about specific code quality issues: 'Let's audit this codebase for unreachable code' or 'This function makes my eyes bleed'
• Have discussions until actionable items emerge
• Explain what should be done and continue discussing until 'I stop having more thoughts to give and the machine stops saying stupid things which need correcting'
• Use 'Ask mode' to walk through examples, share reasoning, and correct the AI when it's wrong

Specific workflow example

Cohen provides a concrete example of how to work with AI on code cleanup:

"There's a lot of things that are both agents and tools. Let's go through and make a list of all of them, look at some examples, and I'll tell you which should be agents and which should be tools. We'll have a discussion and figure out the general guidelines. Then we'll audit the entire set, figure out which category each one belongs in, port the ones that are in the wrong type, and for the ones that are both, read through both versions and consolidate them into one document with the best of both."

He notes that after enough back-and-forth discussion, the AI can often 'do what looks like one-shotting a task,' but emphasizes this isn't truly one-shotting—there's significant human guidance beforehand that clarifies edge cases and potential issues.

Core argument

Cohen's main point is that 'bad software is a choice you make.' He argues that while AI can help clean up technical debt quickly (sometimes in weeks rather than years), developers still need to engage with the code and provide specific guidance. The AI is 'very bad at spontaneously noticing, "I've got a lot of spaghetti code here, I should clean it up,"' but effective when given clear direction.