Boris Cherny, the creator of Anthropic’s Claude Code, recently discussed the current capabilities and limitations of AI-driven "vibe coding"—the practice of building software using natural language prompts. While acknowledging the rapid advancement of AI tools, Cherny emphasized that they are not yet a universal solution for professional software development.
The Distinction Between Prototypes and Maintainable Code
Cherny argues that while vibe coding is excellent for generating "throwaway code," prototypes, and non-critical tasks, it often falls short when producing maintainable software. For critical systems, developers need to be thoughtful about every line of code to ensure long-term viability and stability, a nuance that current AI models often miss.
A Hybrid Workflow
Rather than relying solely on AI generation, Cherny suggests a paired approach for critical coding tasks. His preferred workflow involves:
- Asking the model to generate a high-level plan.
- Iterating on the implementation in small steps.
- Hand-writing code for areas requiring strong technical opinions or specific architectural decisions.
The Current State of AI Coding
Despite the excitement surrounding tools like Claude Code, Cherny admits that models are effectively still "not great at coding." However, he notes that the technology is improving rapidly, moving from simple autocomplete functions to complex code generation in just a year. Industry leaders, including Google CEO Sundar Pichai, echo these sentiments, noting that while AI makes coding accessible and enjoyable, it requires strict oversight for large codebases where security and structure are paramount.
Mentoring question
How are you distinguishing between tasks that require the speed of AI generation versus those that demand the precision of manual coding to ensure long-term maintainability?