Input & process
Everything in the vault arrived through a named channel, never an open inbox that has to guess what something is. What comes in, and where it ends up:
- Articles, podcasts and videos: a share-sheet action on my phone, filed into the external content library.
- Tasks, ideas, journal entries and taste (somewhere I ate, something I read): phone voice and text shortcuts, each landing in its own lane.
- Meetings and lectures: recorded audio, transcribed and routed.
- Course materials and assignments: a scraper logs into the university's learning platform on a schedule, diffs against what it saw last time, and pulls anything new, with lecture slides going into the right course folder and assignments straight into my task system.
- Email: its own triage system, described on the mail page.
The channel something arrives through tells the system what kind of thing it is, so nothing has to be classified by reading it first.
Nothing is processed in place, everything is staged. A voice capture keeps three files - the raw audio, the verbatim transcript saved separately, and the processed note in its final home, linked back to both.
Anything spoken at length runs through one shared pipeline. Meetings, lectures and longer voice recordings all go the same way. A dedicated recorder captures the audio, a cloud transcription service (AssemblyAI) plus Claude turn it into an enriched transcript, and it ends up as a structured note with the people linked and the action items pulled out as a checklist.
Content from the outside world goes through a scheduled sweep. A background job checks the capture inbox every ten minutes, and when there's something waiting it fetches the source, enriches it (title, format, outlet, topic tags, and a check for whether I already have this piece), evaluates whether it's worth keeping, and files the result as a structured note in the content library.
Nothing fails silently. Every lane - the content sweep, the voice pipelines, the transcript router - is built to surface a failure rather than swallow it, whether that's a stuck transcription, an ambiguous lecture-versus-meeting call, or a source that only yields a partial result. Those land on one page I check, not buried in a log I'd have to think to open.