Claude Code Creator Boris Cherny on Vibe Coding’s Limits
▼ Summary
– Boris Cherny, creator of Claude Code, states that AI is still not great at producing maintainable code and is best used for throwaway code or prototypes.
– For critical tasks, Cherny uses a collaborative, iterative process with AI, generating a plan first and then refining the implementation in small steps.
– Despite current limitations, the improvement in AI coding tools over the past year has been dramatic, moving far beyond simple autocomplete.
– Industry leaders like Google’s CEO note that AI-assisted “vibe coding” is making software development more accessible and enjoyable, even for non-technical people.
– However, there are acknowledged limits, as AI-generated code can contain errors, be verbose, or lack proper structure, especially for large, secure codebases.
While AI coding assistants have revolutionized how software is built, enabling rapid prototyping and empowering non-technical creators, their ability to produce robust, maintainable code for critical systems remains a significant limitation. This insight comes directly from Boris Cherny, the engineer behind Anthropic’s influential Claude Code tool. He acknowledges that what’s often called “vibe coding”, using natural language prompts to generate code, excels for disposable scripts and initial prototypes. However, for the core infrastructure of applications where long-term stability, security, and clarity are paramount, a more thoughtful, human-guided approach is still essential.
Cherny explained his own workflow on a recent podcast, noting that for serious development, he uses AI as a collaborative partner rather than a replacement. He typically begins by asking a model to draft a plan, then iterates through the implementation in careful, incremental steps, frequently instructing the AI to refactor or clean up the code. For sections of a system where he holds strong technical opinions or that require precise architecture, he still prefers to write the code manually. This hybrid method leverages AI for speed and ideation while reserving human expertise for the most crucial decisions.
The creator of Claude Code is candid about the current state of the technology, stating the models are “not great at coding” but emphasizing the breakneck pace of improvement. He finds it “insane” to compare today’s sophisticated assistants to the basic autocomplete features available just a year ago. The trajectory suggests these tools will only become more capable, yet for now, they work best within clear boundaries. Top services like Cursor and Augment, which run on Anthropic’s models, and even Meta’s internal coding assistant, demonstrate widespread adoption, but the focus often remains on acceleration rather than flawless final products.
This perspective is echoed at the highest levels of the industry. Google CEO Sundar Pichai has celebrated how vibe coding makes software development more enjoyable and accessible, allowing people without traditional engineering backgrounds to build simple applications. He revealed that AI is now writing over 30% of new code at Google, a figure that continues to climb. Andrew Ng, founder of Google Brain, has also praised the fantastic speed these tools enable, sometimes allowing developers to work while “barely looking at the code.”
Despite the enthusiasm, leaders consistently highlight the technology’s limits. AI-generated code can contain subtle bugs, be inefficiently verbose, or lack the elegant structure required for large, complex codebases. Pichai himself has cautioned that his positive experiences are not in massive, mission-critical systems where security and precision are non-negotiable. The consensus among builders and executives is clear: AI coding assistants are transformative for productivity and democratization, but human oversight and deep technical skill remain irreplaceable for creating software that is truly built to last.
(Source: Business Insider)
