The Canvas pipeline you have already seen is one small piece. This is the shape of the whole thing - how everything I read, record, write and do ends up as files in one folder, and how Claude works inside that folder rather than in chats.
Nothing important lives in a chat. Everything is a plain text file in one folder, and Claude runs inside that folder with permission to read and write.
Obsidian is just a nice editor over a folder of markdown files on my Mac. Claude Code is a terminal that runs in the same folder. The chats are disposable, and the files are the system. Prompts, notes, people, tasks, my own goals - all the same kind of object, a text file with a small metadata block on top.
A sixth layer sits over all of them - rules that check themselves, so the system does not rot.
Every pipeline shares one rule - capture makes no filing decisions in the moment. Deciding where something belongs is the chore that kills note systems, so here the machinery decides later, on its own.
Voice notes from my phone or watch (journal, ideas, tasks, meetings and lectures each have their own lane, sorted by what I said or when I said it). Any link I share from my phone goes to an inbox file that a scheduled sweep fetches, summarises and files. Canvas slides and assignments arrive every morning. Emails with people I track get logged against them.
Press record, share a link, or type a line into a to-do. Everything else lands as a tagged file without me touching it.
The heart of it, and deliberately the most boring part. Folders are physical storage only - notes by kind (my writing, other people's content, meetings, people, companies), never by topic. What a note is about lives in its metadata, so one note can belong to my network, a startup idea and an INSEAD course at once.
Every file starts with a small block saying what type it is, which people and companies it involves, and which areas of my life it touches - from a fixed vocabulary, stamped by the pipeline that created it. A hook checks every write against the schema, so the vocabulary cannot drift.
Because everything is in one folder, every new chat opens onto the same state. A one-page file at the root tells Claude how my world works, and it loads at the start of every session. Cross-chat memory is not a feature I am waiting for, it is what you get from keeping the system in files.
Every view is a saved query over the metadata, computed fresh each time. Change a note's tags and every page that should show it just does.
Everyone in my network has a page carrying one query, show me every note stamped with this person. Meetings, emails, tasks, mentions all appear there on their own. I never update a CRM. I stamp names at creation and the query does the rest.
Every note ends with a block ranking its neighbours by shared people, tags and areas. A conversation from six months ago resurfaces exactly when I open something related to it.
Everything that acts on the store. About fifty skills, each a named text file holding a procedure that runs when I type its name or on a schedule - process this transcript, triage my tasks, write this in my voice. Six consultants, each a one-page behaviour file plus a folder of source material it must cite verbatim - a negotiation coach built on my INSEAD course, a rationality one on the LessWrong Sequences, Paul Graham's essays for startup decisions, and so on.
Seven short files in my own words - who I am, what I am aiming at, my beliefs, my current bets, what I have learnt. Claude reads them at the start of every session and is under instruction to challenge me against them rather than agree with me. They change only when I say so.
A background job runs every half hour, reads new tasks, does the ones that only need vault work and closes them with a note saying so, preps the ones that need me (drafts the email, gathers the research) and leaves anything ambiguous alone. Each morning I get one briefing of what it did and what needs me.
A daily briefing compiled each morning. Documents rendered into a house style from a short description - this page came out of that, as did the Canvas one. Slide decks, small websites, a phone app for tomorrow's readings. The markdown stays the master copy and the polished thing is a view of it.
Anything written to be read by someone else runs through a skill built from my own unassisted writing, with a mechanical check for the tells that make text read as machine-written. It is a floor, not a pass mark, and I still edit.
Every file write is validated, a periodic sweep catches drift, pipeline failures land on one control page instead of failing quietly, and the folder is backed up to git every fifteen minutes. The difference between a system that lasts six weeks and one that lasts years is whether the rules check themselves.
Say I record a 45-minute call with a founder. One skill run later, with no filing by me:
One recording, one skill run, four destinations, and a query that keeps it stitched to everything related.
Prompts, memory, people, tasks and notes are all the same thing - text files in one folder that Claude can read and write. Get the information out of chats and into files, and most of the hard problems dissolve.
It is much less impressive live, which is rather the point. Happy to show you.