One of the more prominent (and disputed) use cases of LLMs has been generating code. Computer scientist Andrej Karpathy of OpenAI (the folks behind ChatGPT) called it Vibe Coding. Simply put it means to use prompts in English to describe the functionality and objective of the solution and allow the LLM to generate the required code. This code is then pushed to production and one gets a working app, website or a tool which can then be hosted locally (on your system) or on the cloud.
When I first tried my hand at it, I felt like I had unlocked a new world full of possibilities. For instance, I used Claude (by Anthropic, Free Version) to create a to-do tracker. I gave all my requirements in form of a short prompt (see below). It understood them and generated a simple code for me.
All I had to do then was save it on a notepad as an html file and then open it on my browser to access it. The result was beyond my expectations. It was fast, accurate and I had gone from an idea to deploying it live within a few minutes.
I was happy with the outcome. So I had to build something more. I tried to build what every HR BP in the world uses at some point in their life – a pulse survey but a cooler one. I have always struggled with the thecomplex dashboards that come with pulse check tools, so with a few prompts it was able to create an easy to understand dashboard as well. Now if I were to deploy it with the help of a few engineers, I would be able to show it to them a live prototype.
Fodder for the Curious
I have always wanted to learn coding. But I always struggled with the motivation and the discipline that one requires to master any skill. I even tried to follow a step by step Python class on Udemy with a friend, but I couldn’t keep up. In the end, I gave up. Until came AI and its fascinating capabilities to create solutions – apps, websites and what not.
However, the software engineering world remains divided on AI assisted coding. While many have argued that AI can actually improve code and improve code quality, relying entirely on vibe coding without understanding the fundamentals of architecture is perceived as dangerous.
There are concerns about scalability and security which make most experts wary of relying only on vibe coding. Some redditors have called it lazy coding for those who don’t understand the heads or tails of coding fundamentas. This describes me quite well. But for a generally curious person, vibe coding opens up new areas of exploration and eventually discovery.
What Next?
AI is transforming our world. It is indeed possible that we might become lazier by the day and start relying only on these tools heavily. But it is not a simple lift and shift as many pundits have spoken. When I used the same prompt of the To-Do list on ChatGPT, the outcome was much different. The more variations I suggested to it, the more it started to mess up the original functionalities. An export to excel button got added but the ability to add new task was suspended.
What did a clean UI really mean? Looks are often subjective so I had to keep refreshing the prompt but things kept breaking. I ran through multiple iterations in order to move beyond the MVP stage. Of course, the paid versions of all LLMs and other no code platforms like Vercel, Cursor, Lovable are touted to be much better and have better functionality but they still don’t rival an open source platform built by someone who knows coding grounds up.
Vibe coding may not entirely replace coding but it could definitely amp up productivity in the short term. Even Ycombinator reported that 25 % of its Winter batch participants were using 95 % of AI assisted tools to develop code. Faster developmental cycles are now possible but the error rate still remains a challenge in the code developed entirely by AI.
In any case, for non-techies like me, using free AI tools to create simpler “tools for one” has been a delight so far!