My take on Vibe Coding

I think Vibe Coding is the greatest threat to systems and the biggest blessing to developers.

If a non-programmer has ideas that they can try to implement, Vibe Coding is the easiest way of explaining it to someone and seeing how poorly explained it was. Failure to communicate will no longer cause two people to exasperate but will make one refine the requirements.

After implementation, it will still have bugs and security flaws. Human developers will be needed to fix it. Some will take the clear requirements and do it manually because it is faster than finding the bugs made by AI.

But… developers can also Vibe Code. Why? First, to see that they also can’t express what they want. Don’t blame the client, blame the humans. Second, to try some things and even learn new languages with examples.

My experience as a lazy developer

I’ve developed quite a few plugins for WordPress to expand a no longer supported plugin. Expect for the SQL fine tuning, mostly of them were done by vibe coding. I tried the general audience choices and between ChatGPT and Gemini, Google’s tool was clearly the best for my needs.

When I had a clear idea of what I wanted, I decided to combine them into a plugin of my own and drop the one I had before.

I tried to write all the requirements (a bit more than 2 A4 pages), attached my plugins as reference, and off we go. This went on for three days to renew credits and go further.

Claude – gave what I wanted quite fast and clear. Was easy to change and add things. The final product was quite solid even though the tokens were ending very fast. Was unable to integrate the plugins, I did it by hand. The end result had around 50kb and I have no idea of what it contains.

ChatGPT – started making never ending questions to align with my desire. Always giving ideas and asking what I wanted next. Run out of credits multiple times, still don’t have anything to show.

Gemini – was close to what I wanted (I hoped so) but had more errors than Claude. There was a time when any change would spend all the tokens at once.

OpenGravity – did part of what I wanted. Still worse than Claude and the final product visually was much less appealing.

Key Takeaways

• AI can be a great tool to speed up your development.
• The sooner your requirements are clear, the faster you start and less tokens you spend in the back and forth. It can also mean “throw that version away and restart with the improved requirements after discussion”.
• You can do it with free tools.
• You still lose notion of what you “did” as you don’t get involved in each line of the code.

In the end, you are still responsible for the bugs and security flaws, so give it a big review. My advise is to ask AI to also audit and give you technical documentation that you can use to validate step by step.

Or just ask a different AI to do it for you. One AI may not take your job, but strength is in numbers.