The Working AI Stack
Every tool that survived six months of real use, what each one is for, and the four that left.
One model per job
Switching models constantly means never learning any of them. Pick one for long reasoning, one for fast drafting, and stop reading release notes as if they were work.
The measurable difference between frontier models is smaller than the difference between a good prompt file and an improvised one.
The gap between models is smaller than the gap between a good prompt file and an improvised one.
What stayed
The survivors have one property in common: they do not own your data in a format you cannot leave with. Plain files, exportable boards, self-hostable flows.
Anything that could not be replaced in an afternoon eventually became a liability, regardless of how good it felt in week one.
What was dropped
Four tools left the stack: two because they duplicated something I already trusted, one because it hid state I needed to see, and one because it only worked when I was already organised.
A tool that requires you to be disciplined before it helps is not a tool, it is a test.