learn-projectlisted

# learn-project > If the conversation is in Chinese, read `SKILL.zh-CN.md` instead — same procedure, Chinese templates. A repository has a **main line** (the path every request takes from entry to output) and a few **branch lines** (families of pluggable members that join the main line at a hub). Later changes rarely touch these; they are what a newcomer needs and what nobody writes down. This skill makes an agent extract them **from code and git history**, with counted evidence, hands the owner a short list of questions, and stores the confirmed result **inside the repo** so it survives closed windows, new sessions and model swaps. Afterwards a task reads 1–2 line files (each < 120 lines), not the codebase. Not a memory plugin, not chat compression, not a vector store, not one big `ARCHITECTURE.md` reloaded every time. Plain markdown + shell scripts; works in any agent that loads skills. ## Vocabulary | Term | Meaning | Mechanical signature | |---|---|---| | **Line** | One end-to-end path with a stable design: entry → steps → output, who owns what, how to add one | — | | **Main line** | The spine every instance passes through | singleton (no siblings), high fan-in, on the path from the entry | | **Branch line** | A family of same-shaped members hanging off the main line | ≥5 sibling dirs with a shared shape **+ a hub** (one file referencing most members) | | **Common layer** | Shared utilities — *not* a line | high fan-in, no hub, no shared shape | | **Sample member** |