Boris Cherny, creator of Anthropic Claude Code, addressed concerns regarding the quality of AI-generated software. He noted that developers increasingly receive emails from peers struggling with 'AI junk code' and shifting workflows where engineers act more as reviewers than primary writers.
Cherny outlined two approaches to AI coding: treating AI output as a black box or maintaining rigorous standards. He suggests that for prototypes or disposable code, the black box method is acceptable. However, production code requires higher standards than human-written code to prevent maintenance issues.
To ensure quality, Cherny advocates for automated safety mechanisms, including end-to-end testing, daily fuzz testing, and automated refactoring. He defines the developer's core responsibility as maintaining quality standards, effectively acting as a 'mini CEO' who decides what code enters the system.
When automated processes fail to maintain standards, Cherny recommends deeper intervention in the AI's workflow. Options include forcing the model to rewrite the entire codebase or waiting for more capable future models to handle the task.
