Alejandro Pena's recent blog post struck a chord. While Andrej Karpathy may have initially highlighted vibe coding's utility for "throwaway weekend projects", I, like Alejandro, believe it deserves far more credit. At the very least, its effectiveness for prototyping is undeniable.
I saw it first hand with my first vibe coding project, which happens to be my first development project. A single prompt, crafted in under five minutes, produced enough to convince me that Infinite Timeline was an idea worth pursuing. This rapid validation is a game-changer.
However, the term "vibe coding" can be misleading. It suggests a passive, almost leisurely activity, where one simply observes and has fun. While that is certainly an approach that can bring joy, building something functional requires active engagement.
Personally, I rarely allow the LLM to implement anything directly. Instead, I request instructions, review them meticulously, challenge the proposed solution, and only then implement them myself if I believe the suggestion is likely to work and, critically, unlikely to introduce major issues. In my experience, a "just see stuff, say stuff, run stuff, and copy-paste stuff" mentality is a surefire path to never-ending cycles of bug-fixing, where every change breeds new problems.
Interestingly, the problems Alejandro points out – the risk of low code quality, vulnerabilities, and suboptimal solutions – are not unique to LLM-assisted development. These are challenges familiar to any team of professional developers, even highly paid humans.
As a "vibe coder", I operate with extreme caution, acutely aware of how my lack of deep expertise can lead to trouble when least expected. I remember spending a whole Sunday working on what I thought would be a quick change, only to end up throwing out the whole day's worth of changes.
Out of cautiousness, I'm deliberately avoiding user registration to sidestep the complexities of securely managing user credentials, but I'm not sure this is that different from traditional software development. Some ventures a risk-free, most aren't.
Similarly, I'm running into challenges to build some of the more complex functionalities I have in mind. For example, I'd love to make it easier on my users to create timelines. My coding assistant and I have tried natural language processing, but we haven't been successful yet. We're also exploring using an LLM, also with no success yet. But again, I'm not sure this is that different from traditional software development. Some things come easy, some don't.
Despite these caveats, the overarching takeaway is positive. Thanks to these tools, I am finally building an application. Thanks to other AI tools, some are finally creating music, creating visual art, etc.
Any technology that fosters creation gets my enthusiastic support and these types of instruments are the most exciting aspect of this new era.
Manuel Panizo Vanbossel