What is the right solution?
Small, tightly bounded, well tested libraries. They'd even be better for AI coding as well, as they'd give them a clearer foundation to work off.
I did an analysis of the number of individual technical desions between the relevant sections of NIP-01 and my NIP-01 ish library. The source code makes roughly 6-8 times more technical decisions than the written document. Meaning that for every abstract decision in a spec, a programmer has to make 7 or so additional decisions on average.
If an AI starts with written specs, it has to extrapolate those decisions itself, which introduces error. If starting with a basically a la carte menu of compostable libraries, it can work at higher levels of abstraction in a much safer context.
I did an analysis of the number of individual technical desions between the relevant sections of NIP-01 and my NIP-01 ish library. The source code makes roughly 6-8 times more technical decisions than the written document. Meaning that for every abstract decision in a spec, a programmer has to make 7 or so additional decisions on average.
If an AI starts with written specs, it has to extrapolate those decisions itself, which introduces error. If starting with a basically a la carte menu of compostable libraries, it can work at higher levels of abstraction in a much safer context.
1👍1🤙1
