Claude Code as a workshop: CLAUDE.md, scheduled `claude -p`, hooks and skills

The write-up is less a product announcement than a field report on making Claude Code behave like a stable workspace. The central move is to keep the AI out of the deterministic parts: the author's scheduled job collects data with Python, then wakes the AI once to read that data, and the AI is allowed to write only into the record notebook. Tools are narrowed from the start, so the unattended run cannot post, send or log in. That preference for narrowing what the AI can do, rather than trusting it to behave, recurs in the hooks and skills sections: hooks run outside the model's memory, and skills attach a name to a procedure so it can be recalled by description rather than by exact command. The closing section acknowledges the limits—visual design remained hard to delegate, and a filename collision caused overwrites on same-day runs—which keeps the piece in the register of a working log rather than a polished guide.