Vibecoding this website with AI in 1 hour
Vibecoding has been a term floating around X and the VC world for a few weeks. The concept to vibecoding is to cede all control over to AI when coding. Three years into the AI gold rush, it's still unclear just how much LLMs will shape our future, especially when it comes to the future of work. Will everyday employees truly turn over important tasks to agentic AI? I decided to explore the possiblity with a low stakes project of building my personal website.
Technologies Used
- ChatGPT - Writing a (Product Requirement Document) PRD for my website
- Cursor - For code generation and terminal commands
- Vercel - For deployment
- EmailJS -Form Functionality
Prompting ChatGPT for a PRD
In the spirit of building for speed and simplicity, I used a very basic 0-shot prompt for ChatGPT: "Help me write a PRD detailing a personal website to be hosted about me. I want this website to include: my work experience broken down in a similar fashion as linkedin, a blog, and about me, with a way to contact me. I want to use Vercel and EmailJS."
This looks promising. Now let's just paste the whole thing into Cursor and see what happens.
- Cursor generated the entire project from scratch
- Cursor prompted me if I wanted to use Tailwind CSS
- Cursor installed all dependencies
- Cursor even helped me set up .env file for EmailJS variables
- npm run dev and the website was live in 30 minutes
Key Takeaways
This felt like magic — building was incredibly fast and fun. However, LLMs are not taking away any jobs soon.
- I still needed to understand basics about React and web development
- Code generated by Cursor sometimes came with errors that required debugging
- Debugging can be done with the LLM but required reading of console logs
- Referencing components required writing code by hand. Prompting was difficult
- There are some lines of code that I don't understand
What has changed
The workflow of creating a PRD, and handing it off to a developer can be done by an AI agent. I can see that PRD workflow becomes automated by Jira/Asana/ etc. AI provides consistency across all specs and also retained context.
Conclusion
Vibecoding was a great way to create something low stakes. It isn't going to be a good option for zero-shotting production level code — especially products with complicated logic and state management. However, it can be a pretty incredible game changer if a current code base can be used as multi-shot prompts for a product manager to create very dynamic demos and prototypes. This opens up a whole new world of getting user feedback and testing MVPs without requiring the company to invest in engineering cycles until something is proven.