你觉得Claude Code写代码已经够爽了？试试做运维 OK我怕被运维好友喷但是我觉得比我强多了

YOLO YOLO dangerously skip permissions

By the end of WWDC 2025, I will try something bold. I will rebuild Gika from scratch, using latest coding agents, and see how automated the process can be. Gika has 30K+ LoC, heavily uses latest SwiftUI and UIKit, tons of custom interactive animation #GikaCodex #Day1

I didn't build anything today. I had hours of conversation with ChatGPT to discuss how our app should look on iPad and macOS, but I didn't let Claude Code to write any single line of code. This is worrisome. #GikaCodex

#GikaCodex Day 3 Spent last 12 hours on learning mise and tuist. Great tools. Haven't used advanced features like caching and selective tests, but as a Xcode generator tuist is already superb. Just copied my basic agentic issue attacker on this. Get CI done. Still scaffolding

Tried to use Claude Code to write some collection view for me very disappointing…you can feel it is way dumber compared to working with React I might need to give a very detailed spec. If only I could have o3 in CLI…

So Xcode gets agent support and we can probably reuse ChatGPT Pro subscriptions?! Your dark knights just get supercharged Peter Steinberger

Four turns and ChatGPT hit token limit? Based on the context window, this is probably 4.1. Never really used them, but they do a plan before code generation. Is this ChatGPT or Apple's prompt? #WWDC25

We get 120Hz on SwiftUI for iPhone, at last. Don't know if my appointment in some event helped a little bit. In any case, CPU driven 120Hz animation, we can build something magical now. #WWDC25 gist.github.com/notcome/3dd742…

Waiting for this for literally two years now

Apple didn't do any codebase/SDK as context thing. Third party model providers cannot edit files. Truly disappointing. #WWDC25

我可真无聊

Best prompt: One more mistake, and I will go find your rack and choke you to death. Understood? Works every time.

dilemma: - keep context long enough, and instructions get forgotten - clear context, then up to luck what pattern Claude finds Iterative improvement feels really painful with agents. I shall try micro waterfall next time

This "I didn't find it so it does not exist" issue is driving me crazy What if we could have LLM records its surprising moments, and use that to guide codebase improvement? But how to capture a codebase/API expectations?

丢掉了Claude Code的我工作效率暴涨

LLM用久了确实会导致大脑萎缩，这中间的trade off估计会在未来几年反复出现。Attention就那么多，能放到语言层做掉白不做

How could GPT 4.1 be so slow

So Xcode used Apple Intelligence as its own Apply model?

谁说早起效率高的？想拉人开会都在睡觉……