The Assistant · EP 01 · DESK
The task desk: let an assistant work your to-do list
This is how the assistant in EP 01 takes tasks off my list while I’m out. I bought no new software for it: it runs on my existing Claude subscription, plus a small DeepSeek call for the morning sort. You need a computer that stays on.
The loop
- Sort the list. At 7:45 every morning a job sorts today’s and overdue Todoist tasks into “I can take it”, “I can prep it” and “you”. I get one text. Tasks I label
for-hermesskip this and go straight to a plan. - Plan, read-only. For each task I pick, Claude Code runs headless (
claude -p) with read access only: files, the web, my connectors. It returns a plan as a page on my phone. - Approve. I reply yes, ask for changes, or drop it.
- Do the work in a sandbox. After my yes, a second run does the work inside its own git folder, so everything it writes is committed and reversible. It can’t run commands, push, send email or edit my calendar. Anything outside that comes back as a separate proposal needing its own yes.
- Report. The result goes on the Todoist task as a comment, plus a page and a text. Only I close the task.
Limits: 3 new plans a day.
Real numbers (Oct 22 – Nov 3, 2026)
- 35 tasks closed
- 24 tasks I stopped myself: a plan I didn’t like, or a task that wasn’t worth it
- Nothing runs, posts or spends money without the approve button
Rules that keep it safe
- Planning can only read. Doing can only write inside its own folder.
- Anything public, paid or hard to undo always waits for a human yes.