AI is “vibing”, so should we !!

“Vibe coding” – when this phrase first popped up on my Twitter feed, I brushed it aside thinking this is another one of GenZ fad lingo. However, instead of fading away, the conversations around it kept becoming more and more intense. Finally, this weekend, I decided to deep dive.

What IS vibe coding ?

This is how Andrej Karpathy, Standford professor and former head of AI at Tesla and co-founder, Open AI defines it:

There's a new kind of coding I call "vibe coding", where you fully give in to the vibes, embrace exponentials, and forget that the code even exists. It's possible because the LLMs (e.g. Cursor Composer w Sonnet) are getting too good. Also I just talk to Composer with SuperWhisper…

— Andrej Karpathy (@karpathy) February 2, 2025

Now let me decode this a bit. Gen AI led coding has been there for more than 2 years, what’s so different now ? The answer lies in end to end capability and precision of what new age tools can do. Think about this – till last year, we would prompt an LLM to generate code related to only a section of an application. And also for that to happen, one had to write descriptive prompts explaining what user wants to achieve. An example below:

Here, we prompted GPT 4o to generate a Python script to calculate factorial of a number – and it does a good job of giving back working code. However, the moment our prompt becomes multi topic and complex, it starts to struggle and returns things which either do not incorporate all instructions, or give messy output.

Enter tools such as Cursor AI and Windsurf – the real enablers of vibe coding. They are not only doing away with shortcomings of LLM generated code, but also take it to another level by:

• Allowing users to code an entire application, not only a part of it – without a need for detailed prompting and instructions.
• Choosing the right tech stack for application, all by itself
• Installing missing components on user’s machine, if needed
• Self correcting the code it wrote without user explicitly highlighting errors.
• Identifying interdepencies in components / files and ensuring all dependent parts are updated, wherever applicable.
• Deploying the application on cloud, if asked by user.

All of this sounds like magic, but is not. I tried my hands at creating a fairly simple application. I fed 3 images of an overtime workflow to the tool and it coded the app (frontend, backend and database) end to end in a matter of 30 minutes – it includes writing code, running the app, debugging and fixing the errors – all autonomously.

Below are the inputs that I gave to Windsurf:

And that’s it … This was the input given to the tool and it created a fully functional app. It auto decided what front end should be used, what back end and what database is the best choice for this app. How cool is that !!!!!

Input fed to Windsurf below:

You can see how it has started to create the relevant files and writing code inside every file. It eventually created an entire folder structure, with multiple files as shown below:

Here’s how the application looks like:

This is a small experiment to play around with “Vibe coding” concept. Although what I have built is preliminary and nothing remotely close to how an enterprise app looks like, yet the potential looks immense. Like everything else, these tools will get better and more useful over a period of time. It is exciting to envisage what the art of the possible will be !!!!!