christopher
@christopher
We spent 2 weeks implementing the New Architecture and trying to understand its benefits and limits. That included asking friends at Instagram, Spotify, and other teams what they thought. It wasn't even close to being ready. And worse: unreliable. We pivoted and spent the winter in "deep work" mode learning Swift..
5 replies
4 recasts
53 reactions
christopher
@christopher
That moment was a "bet the farm" kind of moment. We knew that AI struggled hard with Swift/SwiftUI and understanding the layout of the applications. Worse, there are no crypto packages to lean on like viem or @peterferguson.eth's react-native-passkeys. So we just wrote everything ourselves using Swift Crypto. And got way better at prompting. https://github.com/apple/swift-crypto
4 replies
3 recasts
31 reactions
Shaw
@shawmakesmagic
Do you think this would have been tractable with your team size and experience without AI assistance?
1 reply
0 recast
1 reaction
christopher
@christopher
January to about mid-April we did it mostly without AI, mid-April to now, AI is doing most of the structured thinking, but still gotta implement it by hand. There's quite a bit of hallucination and false confidence/sidequesting it will perform. With crypto libraries you can't go too far off the directive.
1 reply
0 recast
0 reaction
Shaw
@shawmakesmagic
Have you tried Claude Code and shoving all your research and structure into a CLAUDE.md? I found this helped dramatically. Or cursor rules. Spend a lot on context but its WAY better than vanilla
1 reply
0 recast
0 reaction
christopher
@christopher
There is an upper bound to adding in markdown files. It's good for one-shot context, but I find that the few-shot memory context is way more performant.
0 reply
0 recast
1 reaction
