linear's new design is cool and all,
but one thing i really enjoyed is that they used claude code to create an internal one-off tool to iterate on the color palette quickly
one thing i've been doing which i think is underrated is vibecoding custom LLM wrappers for specifc tasks
one example is making a CLI to create consitent images based on the reference image + list of prompts
probs could be done with openclaw but i don't like using chat for this
good piece by @CalebJohn24 on selling to agents
it will be more and more obvious over time that if you're invisible to agents, you're invisible to everyone
https://t.co/kwfOEa1Mo5
started vibing simple playgrounds to polish specific UI elements.
here's a demo app that claude oneshotted so that i can fix the ugly "approve"/"deny" buttons
i'm becoming increasingly annoyed by the way Claude tries to advertise itself through my codebase
e.g. adding itself as a co-author to git commits, creating CLAUDE-md by default
this was fun for like 3 days, now it's just slop
while using Claude Code I’ve noticed that often it performs better if I provide *less* context
now I tend to provide less technical details unless it’s a hard technical problem
surprisingly not much discussion on Apple's Foundation Models.
i believe it's the first OS-level API to gain access to free, offline, and private LLM
w/o having to install anything.
here's a great vid to get started.
https://t.co/zSs1pY77TF
tried Codex
- nice for coding on the go
- the output is decent
- i like that i can preview before sending a PR
- feels slower than cursor
- it creates a new PR for each change, gets super messy
most likely not gonna use much