It feels dishonest to me when senior devs push the idea that new devs can skip reading code and just prompt their way to an app.
It's like famous authors who spent their lives reading telling you to avoid reading because machines can write for you.
What they don't tell you is that they have read so much code during their career that they can skim over LLM-generated code and accept it without really engaging with it.
Because when things break, they can jump into the LLM-generated code, READ it, and fix it.
Maybe they are just excited by how far AI-assisted coding has already gone.
I don't know.
Anyway, READ CODE.
You're right. I've now helped over a dozen people debug their vibed thing. Although vibed code often still reads like it's a string of StackOverflow answers pasted together inside a framework example, it did get better versus the first ones I did when people C&P'd from GPT 3.5.
Since I do like debugging, I have been playing with the idea of offering a "fix my vibed app" service. But it can get expensive real quick because of the tons of code spat out by the bots, so there's a bit of a disconnect in cost basis if someone spent a month with an LLM for 250k sats, and I'd at a minimum charge 1M/day. It feels like too premium a service for a low-effort project template.
The value prop is crystal clear.
What kind of apps are people building?
It is, and there is definite demand. Just, not at that price. And when it comes to deep debugging, LLMs are not a time saver versus doing it manually, also not when you know what you're doing, so time will get spent. The difference is that an experienced coder will do in a day what will take a non-coder with an LLM a week or more, with much more precision.
Mostly specific small webapps and data pipelines that do work but are buggy. The grand visions don't really get anywhere, so I don't see them.
The simplest, most bitcoin-adjacent one was a dude that wanted "pretty" QR codes for his LN invoices on his wordpress site; combine png data basically.
Or for example someone wanted to automate away Excel and go from a csv file straight to a html report (that python code was slop soup and was a headache to make sense of, due to too many iterations by Claude)
Wide range of use cases.
Yes, but that's actually fun!