Vibe Coding: On Second Thought . . .

A few months I wrote a post enthusiastically touting the power of vibe coding to enable non-coders like me to create real web applications through natural language prompts. In about five hours and for a total cost of $50, I built and launched a job board (that has actually already driven a few interviews!). I reckoned that this project would have cost me $15K a few months ago, and taken weeks not days.

With this first project "shipped", I set my sights on a more challenging idea: a marketplace to match agencies and potential clients. This project required more sophisticated coding. A few examples:

• The ability to upload structured data about agencies via a CSV uploader
• A survey that collects data about a potential client's needs
• Displaying different data to agencies and clients (so that agencies couldn't poach info about their rival agencies)
• An API to Gemini so that Gemini could analyze the client survey and then find the best match from the structured data about agencies.
• Automated emails that send results to clients and enable them to automatically get connected to matched agencies

Initially, things went smashingly. I knocked off many of the functions in the same time frame as the job board. Check out some of the screenshots below:

But unlike my five hour journey to create a job board, the process of creating this marketplace has probably taken closer to 80 hours. As I added on functionality, I sometimes contradicted previous instructions I had given to Lovable, meaning that a new feature would work but old features would break or disappear outright.

On top of that, Lovable sometimes made mistakes (I sent dozens of prompts to Lovable that basically said "Hey Lovable, what I asked for isn't working, fix it!). And I did run into situations that seemed beyond Lovable's remit. For example, trying to set up any sort of automatically email program ran into some sort of CAN-SPAM issues and ultimately required using a third party tool that would have to act independently of Lovable.

Indeed, I became so frustrated and tired of trying to build some of this functionality that I eventually decided to just abandon it and release a Mechanical Turk version with a lot of the back-end work being done manually by me. For example, rather than a system where a potential client is automatically introduced via email to an agency, I'm going to just send email intros myself for the time being.

To be clear, I don't blame Lovable per se for these issues. I am certain that a coder or an experienced Lovable prompt creator could have solved a lot of these problems in a few minutes. As the adage says, if you think an expert is expensive, wait until you see what an amateur will cost you. Eighty hours of my time is probably worth about $80K (it sounds like a lot, but that's the equivalent of a junior attorney these days). I probably would have been better off farming out most of this work to the experts.

Of course, Lovable and its competition are all going to rapidly improve. Perhaps six months from now, a lot of the errors and dead-ends I encountered will have been eliminated. Assuming that does happen, I'd still tread carefully. The site I built is not intended for mass consumption. I didn't build robust security functionality, I have no idea if it can handle a lot of traffic, and if things break there's a good chance I won't know how to fix them. In other words, I built something that would probably get me fired if I was actually getting paid to build it.

And maybe this is a ray of hope for developers who are watching AI companies launch products that seemingly threaten their livelihoods. Smart developers recognize that the only constant is change. Wix and SquareSpace didn't kill HTML designers, they just made them shift to becoming Wix and SquareSpace experts. Google ad agencies haven't been destroyed by PMAX, but they have pivoted to become PMAX experts. The same will likely happen with vibe coding tools - we will see webdev pros become Lovable pros. There will be plenty of people like me who will realize that the ROI of paying a Lovable expert outperforms a long and tiring DIY effort.

So am I still bullish about vibe coding? Absolutely. The fact that I can even come close to building a web site with back-end functionality is amazing. Would I recommend these tools to non-technical people who want to build applications for their business? If you can quickly build your prototype or minimum viable product yourself, great! But if you are trying to build something that your customers will actually be interacting with, bring in the experts sooner rather than latter!