Structure your context so every session gets better over time, with memory, briefs, and working files that stay close to the tool you use.
Favor a compact root, a strong CLAUDE.md, and explicit command paths so Claude Code can orient fast in a terminal session.
Start with a lean root `CLAUDE.md`, then add modules only after you repeat the same explanation twice. Keep the repo easy for the CLI to inspect and easy for you to verify.
Ignore big skill libraries, hooks, and fancy context routing until your basic plan → execute → verify loop is stable in one real repo.