I was reading some stuff by Michael A. Jackson (Problem Frames Approach) and T.S.E Maibaum (Mathematical Foundations on Software Engineering) because I also had the impression that too much talk around LLM-assisted programming focuses on program text and annotations / documentation. Thinkers like Donald Schön thought about tacit knowledge-in-action and approached this with design philosophy. when looking at LLM-assisted programming, I call this shaded context. imo agent.md files are not requirements / specifications (yet).

as you say, software engineering is not only constructing program texts, its not even only applied math or overly scientific. at least that is my stance. I suspect AI code editors have lots of said tacit knowledge baked in (via the black box itself or its engineers) but we would be better off thinking about this explicitly.