well actually, i wish i could setup some way for a script to open a pr and the app view shows it was a bot using my credentials

like the tailwind guy saying if you are not freaking out, you ngmi. just full on nft craze bs; not a good look for our industry

but I do agree that agents running async in an isolated environment rules. maybe it's a question of having a good UX on handing the work off/steering the agent when it goes into a tangent that you don't want to

I tend to prefer running them on their machines, the "async" coding agents are built on the assumption that AI can one-shot what you want, which is wrong both in terms of AI capabilities, and also that I don't want to spend much time writing complex prompts detailing what I want. I want to explore!

opus 4.5 the goat of building todo apps

“you shouldn’t source shell scripts from LLMs every time you open your shell”, but the counterpoint is searching and reading bash documentation, so it’s impossible to say if it’s bad or not….

one thing that I tried recently and worked very well is “just-in-time dotfiles”. I got a new work laptop and did all the shell/tmux/vim/etc configs by asking claude thins like “render the active k8s cluster and namespace on my tmux status line”, and they nail it every single time

While I understand the hypothesis, it attracts so many AI slop “it’s not just X, it’s Y” posters. Not to mention the insufferable tone that this approach is inevitable.

There’s no such approach! Nobody is shipping software like that (yet). Ship it first, brag about it later.