Vibe Coding Follow-up

I was inspired in writing this post after seeing this article from Stephen Robles being reposted on Mastodon. It is a very interesting read, and certainly resonated with me in some areas, but also I was surprised by the final result.

The TLDR is that AI wasn’t good enough to take an idea and make it into an app, all the while having zero coding experience.

I’m currently writing from my Tako Writer app, after having made a note for the post idea on my Note Tako app! I guess what I’m trying to say is that AI may be good enough to get you there, but the biggest takeaway should be that a lot of time needs to go into learning the AI tools first, before you can think about start building something with it.

I still get the impression that AI is seen as a magic wand, just wave it in the air and something great will appear out of thin air. In my experience AI is still more of a Range Rover engine that breaks down periodically and I’m the car owner always carrying a set of spanners and grease to keep it running when it breaks.

I wanted to take this opportunity to fully layout my “development cycle” with AI, in case it could help others. Another takeaway from Rob’s post was that he started using the ChatGPT Mac app and a lot of frustrations were with the app getting confused which files to edit and which files were active in Xcode and so on.

I take a different approach as outlined below.

The advantage of this approach is that Codex is in full control of the GitHub repo, can navigate, create, edit, delete files as needed, without me having to know which file and where I need to make the changes. Sometimes you need to fix manually some conflicts, but even those are clearly labelled and super easy to manage.

With this cycle, there is a continuous feedback loop in which every iteration can easily build on the previous step. On top of this, the advantage and the true power of this is the ability of Codex to work on multiple concurrent tasks.

And this is truly the power and, in my opinion, the better way to approach AI development. When I started, I used to write long all-encompassing prompts, asking for a lot of changes all over the app. I quickly realise this was a mistake as often Codex would straight up ignore parts of the prompt! Instead, I started writing multiple bite-size prompts and letting the system work in parallel, achieving much better results.

It’s still not a perfect system, but I also had never touched Xcode and Swift before starting this journey a few months back, but after a lot of trial and error I got there. Yes I won’t be winning any awards with these apps, but I use them daily and I’m so happy with them!

So I guess my TLDR (I know this should go at the top): never give up and learn how to proactively swear at your AI of choice while you try to keep it on a straight line!