The creator of Claude code breaks down where vibe coding falls short

Claude Code creator warns about limits of vibe coding

AI coding tools are everywhere right now, but their creator says they still come with serious tradeoffs. Boris Cherny, head of Claude Code at Anthropic, told a recent interview that vibe coding can feel powerful but often falls short when long-term reliability and maintainability are required.

That weakness becomes clear as projects grow or need long-term support.

What vibe coding actually is

Vibe coding usually means letting AI generate large chunks of code from natural language prompts. Developers focus on results rather than reviewing every line closely.

This approach can feel fast and creative. But Cherny says the convenience can hide deeper issues that show up later when code needs changes or fixes.

Best fit for throwaway code

Cherny says vibe coding works best for experiments and prototypes. These are projects that are not mission-critical and may never be maintained long-term or shipped widely.

In those cases, speed matters more than structure. AI can help test ideas quickly without worrying about future upkeep, documentation, or long-term code quality.

Maintainable code still matters

For serious software, Cherny says maintainable code is essential. That means clear logic, readable structure, and understanding every line.

Cherny says today’s models are still not great at coding overall. They can misunderstand context or produce overly complex solutions.

AI-generated code may work at first, but it can be hard to fix or expand later. That can slow teams down and create technical debt.

How Cherny pairs with AI

For important tasks, Cherny pairs with an AI model instead of relying on it fully. He often starts by asking the model to generate a clear plan first.

He then builds the code in small steps, asking the model to clean up, refine, or improve parts along the way as the project evolves.

When humans write the code

When Boris Cherny has strong technical opinions, he prefers to write the code himself. He says certain decisions still need deep human judgment, especially when performance or security is involved.

AI can assist with suggestions and cleanup, but it cannot replace careful thinking about system design and long-term tradeoffs.

Claude Code adoption has grown rapidly since its 2025 rollout

Claude Code launched earlier this year as part of Anthropic’s effort to bring AI deeper into software development workflows. The tool quickly gained attention from professional developers.

It has also attracted non-technical users who want to build software using natural language prompts instead of traditional coding.

Major tools rely on Anthropic

Some popular coding platforms offer Anthropic models as an option or use them for specific features. Services like Cursor and Augment use this technology to help developers generate code faster and reduce repetitive tasks.

Meta also uses Anthropic models inside its own coding assistant, showing how deeply this technology has spread across the industry and influenced modern software workflows.

AI writes real production code

Anthropic CEO Dario Amodei predicted that AI would soon be writing a very large share of code and has since said that many teams at Anthropic are now using Claude to generate a very large fraction of their code, though usage varies by team.

Google CEO Sundar Pichai has said that more than 30 percent of Google’s new code is now generated with AI assistance, up from roughly 25 percent the previous period, according to the company’s public statements.

Vibe coding gains momentum

Vibe coding has quickly gained popularity across the tech world. Leaders say it makes software development feel more accessible, creative, and less intimidating for people without traditional coding backgrounds.

Google CEO Sundar Pichai said the approach allows people with limited technical skills to build simple apps and websites, helping more users turn ideas into working digital projects faster.

Speed sometimes hides problems

AI-generated code can move fast, but it can also include mistakes, hidden bugs, or poor structure. These issues may not appear right away and often surface later during real-world use.

Some leaders warn that AI works best on smaller projects rather than large systems, especially in areas where reliability, security, long-term maintenance, and accountability matter most.

AI coding still has limits

For critical tasks, Boris Cherny says he pairs with an AI model instead of letting it work alone. He often asks the model to plan first. Cherny says current AI models are still not great at coding. He believes there is significant room for improvement.

He then refines the code in small steps, sometimes rewriting parts by hand when strong technical judgment is needed. He added that today’s tools already feel far ahead of where AI coding stood just a year ago.

As AI opens new doors in tech, it’s clear that mastering AI today can protect your career for years to come.

Conclusion on vibe coding limits

Vibe coding can boost speed and creativity, but Cherny says it cannot replace careful engineering. Maintainable code still needs human attention.

The problems become more visible in large codebases, where small mistakes can cause security or maintenance issues. As AI tools improve, knowing when to rely on them may matter more than raw speed.

OpenAI launched Codex AI coding agent in ChatGPT, making the path into software development significantly easier.

What do you think about vibe coding limits? Share your thoughts.

This slideshow was made with AI assistance and human editing.

Don’t forget to follow us for more exclusive content right here on MSN.

Read More From This Brand:

• AI Tools Taking Over Coding: A New Era for Developers?
• Opera’s AI Wants to Be Your Coding Assistant
• a16z finds startups spending on vibe coding