Claude Wrote the Code for Cloudflare, Developer Reveals Prompts

It is no surprise that developers are using AI models to write their code. However, a developer from Cloudflare sparked discussions when he used Claude to build something, while documenting the entire prompt process. 

Kenton Varda, principal engineer for Cloudflare Workers, revealed that an open-source OAuth library published under the Cloudflare Workers project was largely authored using Claude, Anthropic’s AI model. Varda highlighted that every line of code was thoroughly security-reviewed and met industry standards, ruling out any suggestion that the project was vibe coded.

“I was an AI sceptic. I thought LLMs were glorified Markov chains...I started this project on a lark...and then…the code actually looked pretty good,” Varda wrote in the GitHub repository. 

“Claude’s output was thoroughly reviewed by Cloudflare engineers with careful attention paid to security and compliance with standards,” he added.

The project includes complete prompt histories in the git commit logs, which is an unusual but transparent move. 

Max Mitchell, senior software engineer at tastytrade, mentioned that it offers rare visibility into human-AI pair programming. 

“What’s fascinating isn’t just the code—it’s that they documented every single prompt in their git commits, creating an archaeological record of human-AI collaboration,” Mitchell wrote on X. “AI handles mechanical implementation while humans provide direction, context, and judgment.” 

Developers on Hacker News echoed the sentiment. One comment pointed to a commit where Varda challenged Claude over a design flaw to improve security.

“I don’t think a non-expert would even know what this means, let alone spot the issue,” the user wrote, underscoring the importance of human oversight in AI-assisted development.

The library marks a significant case study in practical LLM usage for critical infrastructure code. While concerns remain about generative AI’s reliability, the OAuth project demonstrates how structured workflows and expert review can transform LLM usage.