claude code vibe-coded a swiftUI markdown-first notes(.)app replacement in like 20 min unattended.

Im truly like wtf

I’ve heard / experienced a bit that AI for swift is bad because of their poor documentation and lack of open source projects! Very curious about how well it works

agentic

@stackdotso. Prev CTO/cofounder of mirror. strangechances

i think context beats training for such kind of stuff, more so nowadays with 200k-1M token windows.

Most of the times i can just provide good initial sources so making a -modification- on top of existing code is idiomatic and follows the same style.

One shotting is hard, but agents really gain from the same feedback loops human coders use, but a break-neck speed.

Ai Agents can use cli's like wizards with how cheap tokens are

Yes that makes sense, though I would imagine it’s counteracted by how much open source JS and Python there is to train on! It would be interesting to see a benchmark across languages for how often LLMs can one-shot solutions to various problems - which would entail both training plus guard rails

This requires agentic ai with search capabilities, at first try it will hallucinate bad deps, then i tell it to -research SOTA dependencies/current ones- and it just finds whatever is best/decides whats better maintained and uses taht instead.

then it can also fetch github (either docs, or code) or peek on node_modules equivalent build deps to see the actual source code

another trick is to tell it it can use the terminal tools to do better digging'refactoring than editing files whole-sale.

Im really digging llms'/agentic ai with swift/go/rust due to the tighter feedback loops its more terse type and safer system rails provide versus say js or python