The more I use AI the less I understand the fascination with “one-shotting.”
Nothing worth doing is one-shottable. It’s some sort of silly teenage sniper fantasy. Cheap stunts at best. Or you’re just starting with something that someone has already put hundreds of shots into, like a famous and carefully formulated unproven math theorem with bookshelves worth of context worked out over decades.
Getting the surprising amount of detail of reality to mesh with the surprising amount of detail of intentionality involves working through a lot of friction that moves at the speed of actual context discovery, not at the speed of either machine inference or human flights of fantasy. And it’s very satisfying to iteratively get it right.
Getting the “context files” right isn’t a configuration preliminary. It’s the whole project. When you get the context setup right the project will be done.
I’ve written 2 posts that I feel get at the heart of this matter. Gonna think about this a lot more.