As AI reshapes software development, Boris Cherny, the creator of Anthropic's Claude Code, believes developers' primary responsibility is not writing every line but upholding code quality. In a recent email exchange, a developer asked Cherny how to handle "AI junk code," a query he says he receives daily.
Two Approaches to AI-Assisted Coding
The developer outlined two contrasting methods. The first involves writing code as before but using AI to speed up the process, ensuring that code remains reviewable and maintainable. The second, more permissive approach, treats code as a black box, focusing only on output. Cherny responded that both have their place.
For prototypes and one-off code, Cherny says treating them as black boxes is acceptable, since they will be discarded and failures have limited impact. However, production code demands higher standards. He notes that Anthropic has implemented extensive safeguards, including coding standards checks, tests, Claude-driven end-to-end tests, daily fuzzing, automated code review, and refactoring.
Without such measures, code can degrade into an unmaintainable mess. Cherny suggests that as models improve, these tasks become easier, such as running routine workflows or using Claude Code Review. He advises developers to intervene directly if issues persist, have Claude rewrite the entire codebase, or wait for the next model.
Cherny describes the new role as a "mini CEO," where developers act as managers, deciding what code enters the system and what gets rejected. According to IThome, this shift reflects a broader change in engineering roles, with some embracing the convenience while others find the work isolating.
Based on reporting by IT之家. Edited and published in English by geisou.
