Boris Cherny's Experiment to Rewrite the Claude App Using Claude Code
Attempting to have AI rebuild its own application in Swift for performance improvements.
1The Experiment's Inception
At Y Combinator's Startup School 2026, Boris Cherny shared his experiment pushing Claude Code to its limits. He tasked the AI with rewriting Claude's Electron-based desktop app in Swift, emphasizing the crucial role of verification mechanisms in AI-assisted development. Cherny highlighted that most teams overlook proper validation processes when working with AI tools.
2Technical Execution
Cherny first set up a Mac virtual machine on GitHub to give Claude access to the necessary environment. He then created an empty Swift project and instructed Claude to rewrite the Electron app. The AI was told to run the original app, take screenshots, and compare them pixel-by-pixel with its Swift version. This process has been running continuously for over two weeks without completion.
3Current Application State
While Cherny claimed the Electron-based Claude app transitioned from 'sluggish and unreliable' six months ago to 'pretty awesome' today, independent tests reveal significant issues. The app takes 30 seconds to launch on Mac, with the last 15 seconds showing the spinning beach ball cursor. This performance contradicts Anthropic's official statements about the app's quality.
4Fundamental Design Flaws
Experts note that Claude's core issues extend beyond technical implementation to fundamental design problems. The application ignores Mac UI conventions and delivers a consistently poor cross-platform experience. Simply rewriting the app in Swift without addressing these design flaws would only improve technical performance while maintaining the unsatisfactory user experience.
5Limitations of AI Approach
Cherny's experiment highlights boundaries in AI-assisted software development. Having Claude attempt pixel-perfect recreation of the existing app demonstrates AI's current inability to generate creative design solutions. This case study reveals important limitations in using generative AI for complete application rewrites.
6Industry Perspective
Application performance remains critical for AI companies like Anthropic. While cross-platform solutions like Electron enable rapid development, they often fail to match native app performance. The industry continues to grapple with these trade-offs, exploring various approaches to balance development speed with user experience quality.
7User Experience Expectations
Modern users expect applications to deliver both functionality and seamless performance. Lag and freezing issues in AI tools like Claude directly impact user satisfaction. Native app performance has become a key differentiator, especially for professional users who rely on these tools for daily productivity.
8Future Implications
Cherny's experiment provides valuable insights into the potential and limitations of AI-assisted software development. While AI will increasingly contribute to code rewriting processes, human creativity remains essential for design and user experience innovation. This case study offers important lessons about identifying appropriate applications for generative AI in technology development.
Original article: Daring Fireball — Claude Code ile Uygulama Yeniden Yazma Deneyi