Creator of Claude Code advises against overusing "vibe coding"

Vibe coding is useful but not all-powerful or universal, Anthropic engineer Boris Cherny — who helped develop the AI assistant — said in a recent episode of The Peterman Podcast.

Leave a comment

Cherny notes that vibe coding is appropriate when writing prototypes and temporary code: «I often do it this way. But you can’t always. Sometimes you need code that’s easy to maintain. Sometimes you need to think carefully about each line,» he said.

Claude Code was launched earlier this year. With Claude Code, Anthropic aims to expand AI adoption in code development processes. But it has also become popular among people without technical backgrounds who want to create simple things using prompts. Anthropic CEO Dario Amodei stated in October that Claude Code writes 90% of the company’s code.

Cherny says that for critical tasks he usually partners with a model to write code: he first asks it to generate a plan, then implements that plan through small iterations — asking the model to improve or clean up the code along the way.

However, he writes components with strict requirements entirely on his own, saying that models «don’t code very well» yet.

«They still have room to grow. What we have now is the worst they’ll ever be going forward,» Cherny says. But he acknowledges that current models are vastly better than what was available a year ago, when AI-assisted coding was mostly just advanced autocompletion.


Читать на dev.by