Using Claude Code for three months, the directory structure has become more and more complex, with rules/ docs/ memory/ skills/ and various layers, but I kept struggling to clearly explain what I was building.

Until I read this paper "Everything is Context," which translated my folder structure into academic language:
→ The paper calls it Scratchpad (Temporary Workspace)
→ Fact Memory (Project-level Fact Memory)
→ Experiential Memory (Cross-project Experience)
rules/ auto-loading vs docs/ on-demand loading → Context Constructor (Selective loading within token budget)
The part that resonated most with me was the section on token window constraints. I used to load rules/ in full, and the context would explode. Later, I split it into two layers to stabilize. The paper calls this "bounded reasoning capacity"—turns out, I was solving this very problem.
Practice comes first, naming comes later. First, stumble, then read the paper, and realize that all the pitfalls I encountered have names.
Original paper text