Fixing your vibe coded mess

Bleh. Late 90s I was a mobile pc tech and I could have used these gloves back then on some computers. People are gross.

I wouldn’t want any other person to touch my keyboard. I blame the lack of accessible cleaning supplies and the fact that if anyone higher up sees me not staring at my screen with a hand on the keyboard they’ll think I’m lazy and not working. Which is true, but they don’t need to be thinking that.

If you redlined a vehicle all the time it’ll break down faster.

LLM’s will do as good code as the average coder. The problem with that is the average coder isn’t very good at coding.

We currently have a semi-serious project at $DAYJOB, like we’re basically allowed to work on it as a team building thing. And one guy who’s tugging along has ten years more programming experience than me, but no experience with the programming language we’re using, so he’s been generating everything with LLMs.

He knows to write unit tests and well, the programming language in question is Rust, which’s strict compiler prevents lots of bad code from happening. So, this isn’t your stereotypical videcoding.

But …yeah, it’s still been challenging to work with.

Yesterday, the guy built a feature which basically gives the user instructions how to create a bookmark in their browser. There’s a few ways to implement this:

• You don’t. Our userbase is gonna be technical, they probably know how to do that.
• You show the instructions for all browsers and let the user pick which instructions to follow.
• You ask the user interactively what browser to set up and then show them only the instructions for the chosen browser.

Right, and apparently the fourth way to implement this, which the LLM generated, is to detect what the default browser of the user is.

Leaving aside the problem that some users will want to set up different browsers than their default browser, how do you implement that? Is there some nice, cross-platform API for it? Well, if there is, the LLM didn’t know about it.
And neither are there nice APIs per operating system. On macOS and Linux, it runs some random commands to access this information. On Windows, the generated code looks at the Registry.

All of this is absolutely horrid to maintain. I do not want to be testing on each OS separately. I do not want hundreds of lines of code for a feature that’s not actually needed. And the worst part is, the guy should know this. He has the experience.
But I’ve seen the guy when he chats with an LLM, just falls into an absolute trance. Does not surprise me that he’s unable to take a step back to think, if this even makes sense to do…

I would to now where I slot in on that graph to know if I can agree with you or not.

detect what the default browser of the user is

Please don’t do stuff like this. You will always get it wrong. I hate apps trying to be “clever” placing unneccessary friction upon my workflow, because they always get it wrong. Then I have to figure out how their regarded auto detection works so I can trick it into doing the rigjt thing.

The clankers always try to be clever.

I do agree, yeah. If we would have auto-detection, I’d want a manual override to be implemented for sure.
It’s a CLI, so it wouldn’t be too bad to add a --browser=firefox flag or similar, but it just makes it even more obvious that this feature only adds complexity, because we would have two solutions when we hardly needed one.

I vibe coded a 2000 line python program. It runs and does what I want. So many weird things in the code. But I don’t care. This code won’t be used to launch a space shuttle or even be used by anyone but me. It works…send it.