The task system
Capturing something takes about five seconds, from wherever I am. I speak into my phone or type a line, and it lands straight in the inbox. Voice capture is transcribed by a local Whisper model on my own Mac rather than sent to a transcription API, so nothing I say into it reaches a third party before I see it written down. From my laptop I can write a task in a sentence and it gets classified and filed the same way. Both phone routes also drop a calendar hold and a silent reminder as backup, so a capture never depends on one system alone.
Every morning a triage pass looks at everything so I do not have to. It reads new items in context, merges anything that is really a duplicate of something already tracked, scores whatever arrived unscored, flags projects that have gone quiet without a deadline, and archives what has been closed out. What is left is a clean queue, not a pile to sort.
Every open task carries a priority score, urgency multiplied by importance, so the working queue ranks itself rather than me re-deciding the order every day. A real deadline pulls a task up the list as it approaches; nothing gets a deadline that was not actually said or written somewhere first.
Two surfaces show the state of play. The Dashboard opens on the week ahead, every task with a day assignment grouped by day, today first, then one tabbed block holding the active queue, the parked list, what arrived in the last fortnight, and everything. A second, narrower page holds just today: today's and overdue tasks, closeable in place, next to today's mail and reading. Day assignments are written three ways: a parser resolves an unambiguous date hint at capture time, the resident agent resolves ambiguous ones in context within the half hour, and the rest get assigned at a planning sitting; nothing ever overwrites a date already set.
The system can close tasks itself, but it has to show its work. Every closure carries a record: which route closed it (the Butler did the work, a condition I wrote came true, or evidence turned up that the world had moved past it), when, and what the evidence was. It is a claim for me to check, not a verdict. Closed-by-Claude tasks sit in their own view rather than disappearing; agreeing is one status flip, and overturning reopens the task and logs the disagreement, so the system learns where its judgement was wrong. If Claude did the work itself and I reopen it, I am asked whether the work should be undone too. And the system errs towards leaving things open: a wrongly-closed task costs me a moment noticing, a wrongly-open one costs nothing.
It is delegation with a paper trail, not delegation on faith.