Building 30+ SDKs in 3 Days: An AI-Powered Development Sprint

Hello Everyone! I'm excited to share my take on Cursor AI-assisted coding or you can say vibe coding.
I recently challenged myself to build as many SDKs as I could in 3 days and ended up with over 30 SDKs including Twitter/X, AWS-SES, Discord, YouTube, Slack, GitHub and many others. Rather than using official npm packages, I wanted lightweight alternatives that only included the features I actually needed.
For this rapid development sprint, I used Cursor and provided it with API documentation URLs. The AI would build the initial SDK structure based on the docs, and I'd test the implementation with my own API keys to ensure it actually worked.
What was interesting about this process was my approach to debugging. When Cursor couldn't solve an issue after 3-4 attempts, I would debug by cross-checking the code against the API documentation. Once I figured out the problem, instead of fixing it myself, I'd hint Cursor in the right direction and ask it to implement the solution. This wasn't just about saving time (though I won't deny enjoying that benefit) – I genuinely wanted to test how well Cursor could adapt and learn from specific feedback.
The results were mixed. For many bugs, Cursor could adapt once I identified the problem. But there were notable cases where it kept repeating the same mistakes due to hallucinations - especially with complex OAuth flows and custom encryption schemes. It would confidently implement incorrect solutions despite clear debugging information.
In the end, this project delivered exactly what I wanted: a collection of ultra-lightweight SDKs tailored to my specific needs and valuable insights into the current state of AI coding assistants. They excel at generating standardized code patterns but still need human guidance for complex authentication and security implementations.
As this is my first post, I'm planning to share more of my learning experiences and projects going forward. I'd love to hear from others who are building lightweight alternatives to bloated libraries. What's been your experience?