Once the notes and Claude.ai's memory are in place, driving the thing day to day comes down to a few short words you type into Claude.ai. No dashboard for this part, and no command line. It is all just chat.
The queue
Jobs waiting to run live in a short list - a queue - held in Claude.ai's memory and backed by a BrainMop note, so it survives from one thread to the next. You add jobs to the list; Claude.ai works down it from the top. The queue is the one piece of state you actually watch.
Claude.ai takes the top item, dispatches it as a ccsw block, then removes it from the list.
The handful of prompts you actually type
Nearly all the driving is these three:
qShow me the current queue - what is waiting, and in what order.
add a job to the queue: sync the SleepTracker portfolio and deploy the changed filesClaude.ai appends that as a new job on the queue note. You are just talking; it does the bookkeeping.
autoAutopilot. Work through the whole queue unattended: take the top job, dispatch it, wait for the result, then the next, without stopping to check in. Typing autopilot does the same thing.
How a job leaves the queue
When Claude.ai dispatches a job it pops the top item: it writes the ccsw block for that job and removes that entry from the queue note in the same step. So the queue only ever shows what is still waiting, and if a thread is handed off mid-run the next thread reads an accurate list. The one rule it holds to while popping is the workflow rule from the last section, one job per repo in flight, so if the top job is for a repo that is already busy, it takes the next job that is free instead.