Getting trivial shape sizes wrong, smushing together different versions of the API, inventing methods that don’t exist. It can be kept on rails by feeding back errors, but the inability to get things right first time almost ever is astounding coming from Python.

chatgpt.com/s/t_69374614...

It just straight up misspelled “devtools” as “develtools”. What is even happening!

It’s a bit saccharine in places, GPT5 does a better version, but it’s really approaching splitting hairs territory on this one

Everyone has their own vibe evals, and mine are “be a New York School poet” what of it

Claude out here quoting Phil Karlton at me like a team member trying to dissuade a dumb manager from making them to something ill-advised. (Claude is right to do this)

This skeet is brought to you by me being destroyed by how hard this song goes open.spotify.com/track/5KyOUI...