claude code first thoughts
I dislike AI but after hearing so much hype about Claude Code I just had to try it out. I bought the $20 plan and created a project I've had on my mind for a while: a simple webpage to display parkour gyms and their free shifts in Helsinki area. Here is the end result and here is the code.
damn this thing is smart
First thought was that this thing is pretty smart. Almost any prompt I gave it was able to do in one shot. I kept my prompts small, and the codebase was small and simple. But still, it was able to hack through most of my requests like a good junior dev, just a bit faster.
When it didn't one-shot, it usually got it on the second try after a more detailed prompt.
means to an end
I like good, clean, clever, short code. I think of a codebase as a garden; the start small and grow over time. You plant seeds (code) and it can be beautiful or ugly depending how they're planted, but nevertheless, it requires maintenance. Sometimes you need to pick out old dying code to let the new stuff grow. A good garden is one where you think of every detail with care, not just get it done and call it a day.
When vibe coding with AI models you really need to throw that thinking away. You're not doing art, you're getting shit done. If it works it works. I had to turn off that part of my brain which actually cares about the minuscule details of the code.
you're the architect
It's important to know how programming works and how to build projects to get the most out of the models. My repo was basic html/css/js with no dependencies, and even with that setup it was important to tell the AI where to code, otherwise it would have just kept building a one-file-wonder. Sometimes I told it to just refactor the code to make it cleaner and more readable, sometimes I told it to refactor specific things.
use git
I just gave Claude Code permission to do whatever it wanted every time it asked. No point in manually approving every file it wants to edit, just let it cook and check the changes later in vscode or some other program.
After prompting a change, confirming that it works and reviewing the code (I just skimmed and made sure it passed my vibe check lol) commit, and commit often. Sometimes the AI just decides to be stupid and do something you don't want (or even something you told it to do, but changed your mind, like a proper PO) so it's good to be able to revert to your last commit.
stealing designs taking inspiration
Claude code can handle pictures and urls. One thing that really blew my mind was after finishing the basic layout of the website, I just downloaded an image from dribbble.com, dragged it to the CC prompt and saying "make the website look like this". It actually worked! The UI looked pretty good. That's awesome.
CC can also browse the web so you can just give it a url (like hanki.dev) and tell it to make the app look like that.
conclusion
After finishing 1 project and playing it for around 10 hours, I gotta say that I'm impressed. Impressed enough to make a blog post of it even though I kinda hate AI.
I have some more projects in my mind, some simpler and some a little bit more complex. Most excited to see how complex stuff it can handle.
✌🏼 Like my content? Subscribe via RSS feed.