Boris Cherny, the engineer behind Anthropic’s Claude Code, has issued a cautionary note about the limitations of “vibe coding” despite the explosive popularity of AI-assisted programming tools. Speaking on “The Peterman Podcast” this week, Cherny emphasized that while vibe coding—using natural language prompts to generate code quickly—works well for throwaway prototypes and non-critical applications, it’s “definitely not the thing you want to do all the time.”
Cherny explained that developers still need maintainable, thoughtfully crafted code for production systems. For critical coding tasks, he employs a hybrid approach: pairing with AI models to generate initial plans, then iterating on implementation in small, controlled steps. For system components where he has strong technical opinions, Cherny still writes code manually, acknowledging that current models are “not great at coding” despite rapid improvements.
Claude Code, launched earlier this year as part of Anthropic’s push to integrate AI into development workflows, has gained significant traction. Top AI coding services like Cursor and Augment run on Anthropic’s models, and even Meta uses them in its coding assistant. The tool has particularly resonated with non-technical developers seeking to build software through natural language prompts. Anthropic CEO Dario Amodei revealed in October that Claude was writing 90% of the company’s code.
The broader AI coding revolution is transforming the tech industry. Google CEO Sundar Pichai disclosed that AI now writes over 30% of new code at Google, up from 25% in October 2024. Pichai praised vibe coding for “making coding so much more enjoyable” and enabling people without technical backgrounds to build simple applications. Andrew Ng, founder of Google Brain, noted in May how developers can now write software “barely looking at the code.”
However, industry leaders acknowledge significant limitations. AI-generated code can contain mistakes, be overly verbose, or lack proper structure and security considerations. Cherny emphasized the dramatic progress from just a year ago when AI coding amounted to “little more than type-ahead autocomplete,” calling current capabilities a “completely different world” while noting “this is the worst it’s ever going to be” as models continue improving rapidly.
Key Quotes
I do this all the time, but it’s definitely not the thing you want to do all the time. You want maintainable code sometimes. You want to be very thoughtful about every line sometimes.
Boris Cherny, creator of Claude Code, explaining the limitations of vibe coding and emphasizing that while AI-assisted coding has its place, developers still need to prioritize code quality and maintainability for production systems.
There’s still so much room to improve, and this is the worst it’s ever going to be.
Cherny acknowledging current AI coding models aren’t yet great at coding while expressing optimism about rapid improvement trajectories, suggesting today’s limitations will soon be overcome as models continue advancing.
Things are getting more approachable, it’s getting exciting again, and the amazing thing is, it’s only going to get better.
Google CEO Sundar Pichai describing how vibe coding is democratizing software development, enabling people without technical backgrounds to build applications while making coding more enjoyable for experienced developers.
I’m not working on large codebases where you really have to get it right, the security has to be there.
Pichai acknowledging the limitations of AI-generated code for enterprise-scale applications, particularly regarding security and reliability requirements that demand human oversight and expertise.
Our Take
The nuanced perspective from Claude Code’s creator reveals an important maturity in the AI coding conversation. Rather than the typical hype cycle extremes—either AI will replace all developers or it’s completely useless—Cherny articulates a pragmatic middle ground that practicing engineers will recognize. His hybrid approach of using AI for planning and iteration while maintaining human control over critical components represents the likely near-term future of software development. The dramatic statistics from Google and Anthropic about AI-written code percentages might seem alarming to developers, but Cherny’s emphasis on context matters: much of this is likely boilerplate, prototypes, and non-critical code. The real story isn’t replacement but augmentation, with AI handling routine tasks while humans focus on architecture, security, and maintainability. As models improve, this balance will shift, but the fundamental need for human judgment in production systems remains.
Why This Matters
This story highlights a critical inflection point in software development as AI coding tools transition from experimental novelty to production reality. Cherny’s cautionary perspective matters because it comes from someone building the very tools driving this transformation, lending credibility to concerns about over-reliance on AI-generated code. With major tech companies like Google and Anthropic reporting that AI writes 30-90% of their code, understanding the limitations becomes essential for maintaining code quality, security, and maintainability.
The tension between rapid prototyping through vibe coding and the need for robust, production-grade software reflects broader questions about AI’s role in professional workflows. As non-technical users gain coding capabilities, the industry must balance democratization of software development with quality standards. The acknowledgment that current models are “not great at coding” despite their utility suggests we’re in an early phase of a longer transformation, with significant implications for developer roles, software architecture practices, and technical education.