avatar

Boris Cherny on why AI code needs stricter checks

Tighter controls for production code from tools like Claude

avatar
1 day ago

TL;DR:

  • AI code works okay for quick prototypes but not production
  • Unmanaged AI output can create maintenance and security problems later
  • Anthropic uses linting, tests, fuzzing and reviews to keep things safe

Headline

Boris Cherny thinks AI-written code for production needs more checks than code from people. He points to how Anthropic handles Claude for development.

Summary

Cherny says it's fine to use Claude code as a black box for quick prototypes that aren't risky. But for real production stuff, you need stricter rules or you'll end up with code that's hard to maintain and has security holes down the line. At Anthropic they use a bunch of checks: linting, tests, more tests generated by Claude, fuzzing, auto reviews, security scans, and refactoring.

Analysis

People are realizing that just getting the AI to spit out code isn't the hard part anymore. The real work is fitting these tools into solid engineering processes. Companies will need lots of automated checks, security reviews, and custom rules for their own codebases. Cherny mentions things like Claude's code review feature, a CLAUDE.md file, adjusting how hard the model tries, and using the latest models. It looks like Anthropic wants Claude to be part of a full development system, not just a chat tool. In the industry, the edge might come from how well you integrate and control these tools, not just raw smarts.

Impact Assessment

Significance: Medium Categories: Developer Tools, Industry Trend, Technical Insight