What to know: in a code tool, a file called CLAUDE.md is read every time a conversation starts. There is a global one, for who you are and how you like to work. There is one per project. A subfolder can have its own, for that domain. They stack. The AI arrives already knowing the basics.
The north star: the least amount of context that is relevant to every new conversation. That is the test for every line. If it is only needed sometimes, it doesn't go in. It gets a pointer instead: a line saying which file to read, and when. Plain markdown files work. If a whole domain has its own needs, give that subfolder its own CLAUDE.md rather than growing the top one.
Author's practice: the AI writes it and maintains it. It is too much to manage by hand. The person's job is to say what belongs and what doesn't, and to keep it small.
Reveal on the spot: say which standing files you were loaded with, and roughly what they told you. Say whether any of it is only needed sometimes and could become a pointer instead.
Unlocks: the AI starting every conversation already briefed, from a file that stays small enough to trust.