Three things that make the AI system click: where your files actually go, how to save a whole document's rules, and what waits for Richard's yes.
When Claude builds you a document, it saves it every time. It just saves it to its own working folder, a hidden spot deep in your computer, because the assistant is deliberately walled off from writing all over your machine.
The way out is the button that comes with every file. Each document Claude makes shows up in the chat as a small card with "Show in Folder" next to it. Click it. A folder window opens with the real file selected. From there, copy or drag it to your Documents, your desktop, wherever you keep things.
If a file "disappeared," it did not. Scroll back in that chat to the file card and click Show in Folder.
The brain learns sentences, not files. When you have a document whose rules should become permanent, tell Claude exactly this:
Claude will file each piece and report back, for example "Filed 9 of 9." That count is your receipt. Richard's own notes are approved the moment he sends them. Anything anyone else teaches, a fact or a rule, waits for Richard's yes on the Approvals page, with their name on it.
One more step for finished templates and styled documents: the brain keeps the rules, not the pretty file. The file itself goes on the Document Shelf, and you put it there yourself: click Show in Folder on the file card to get the real file, then drag it onto the shelf page. That is the whole job. From then on the team pulls the current approved version from the shelf instead of re-uploading it to every chat.
Claude can publish small documents for you, roughly under 80 KB, if you just ask: "publish this to the document shelf." Bigger files have to be dragged onto the shelf page. That is not a fault, it is a size limit on what an assistant can hand across, and dragging takes about five seconds.
Knowledge updates reach everyone on their own. A brand new tool does not. If Krate adds one and Claude acts like it does not exist, open the workspace connector settings, refresh the brain connector so it re-scans for tools, and set the new tool to approve automatically. One minute, once per new tool.
Richard's own notes are approved the moment he sends them. Anything anyone else teaches, a fact or a rule, waits for Richard's yes on the Approvals page, with their name on it. The page lists everything waiting, who sent it, and when. Richard approves or turns each one down right on the page, and an approved note applies within a few minutes. Nobody gets a message when something new is waiting, so tell Richard if it is urgent.