← Back to context

Comment by eclipxe

1 month ago

I used to feel that way but my perspective over the last 6 months has changed greatly. You can absolutely build very complex things in this style

That's actually not a contradiction.

As far as I can tell, you can build very complex prototypes. But unless these prototypes can be both trusted and maintained, that's all they are.

  • > As far as I can tell, you can build very complex prototypes.

    I think GP was saying you can do more than prototypes. I agree, but it's not (yet) universal on where you can apply it. The best case for my projects has been in trivial but tedious "3rd party integrations". Say you have a mature product but client x wants integration with product z. We are now at a point where we can say "this is our internal model {json dump}, this is the 3rd party integration docs / example {code dump}, write interfaces for this". And it works most times. For things that are a bit more complicated, /architect first and then "now write it" w/ some things from the architect session in context also works.

    YMMV but don't dismiss it out of habit. Things are moving very fast in this space, and I choose to focus on what works now, not on what doesn't. I'm well aware not everything works, but when it does it saves a lot of time.

I wouldn't trust vibe coding to a junior. This would be a recipe for desaster. It's a skill best paired with a very senior dev who can correctly assess the output of the AI.

show me an example of such a complex thing obtained after just writing prompts

  • Sometimes you enter a codebase and it looks like there's some obscure attempt to summon a Lovecraftian entity made of spaghetti and duct tape. Those codebases are both the shitiest and most complex stuff I've ever seen. They sure don't work well and I'm not even sure do a good job at summoning the old ones. I don't find LLM code significantly differing from that kind of monstrosity. It sure is complex.