Claude Code's Boris Cherny: Black-box AI code OK only for prototypes

The question Cherny answered came from a developer with 12 years of experience, who had been weighing two opposing views: that AI-written code should be fully checked by humans, and that code should be treated as a black box and judged only by its output. The developer's frustration was that generating code with AI is easy, but deciding whether a method the AI chose is appropriate for production takes time. Cherny's reply was that there is room for both.

What separates the two camps in Cherny's answer is purpose and risk, not a single rule. He points to Anthropic's own production setup, where Lint, E2E tests, fuzzers, code review, security review and documentation upkeep act as guardrails, and tells developers their job is to hold the bar on code quality. He also notes that as Claude's models get smarter, quality control mechanisms become easier to run.

The writers' own practice is presented as a small-scale case: for prototypes such as data analysis, he does not read AI-generated code in full, though he does not treat it as a complete black box either. He keeps a sense of what code is being written, and asks AI to redo work when execution feels off. His stated priority is not trusting tests blindly, since tests like expect(true).toBe(true) can pass without meaning anything. Whether the black-box approach holds up likely depends on how far teams extend trustworthy tests and written specs — the two starting points he suggests for small teams — as AI takes on more of the coding work.