Four inboxes, one vault — and an assistant inside it

How my AI OS started: not with an app comparison, but with a folder of plain text files and an assistant that could work in them before I knew the shortcuts.

In January my day started with a checklist: clear the OneNote inbox, the OneDrive inbox, Google Drive, iCloud. Four places where notes, receipts and half-finished thoughts landed, all to be visited before I could think about anything else.

The structure was already there: I had taken Tiago Forte’s PARA method (Projects, Areas, Resources, Archive) from his YouTube videos in December, and by early February the same numbered folders existed in three systems. I did not have an organisation problem. I had a too-many-places problem.

A video, and a file format

The way out came from YouTube again: in late February a long run of Nick Milo’s Linking Your Thinking videos showed me Obsidian. What stuck was not the app but what lies beneath: plain Markdown files in an ordinary folder. No database, no export button, no vendor deciding their fate in five years. Any editor or program can open them, including an AI.

Markup was not new to me. I have worked with computers since the end of the 1970s; in the 1980s, in my first training as a mathematical-technical assistant (Mathematisch-technischer Assistent), I looked at SGML and formatted mainframe reports, and later learnt HTML. New, and only in this decade, was one simple markup language as the everyday format for all my documents — which is also far cheaper for an AI to read than a PDF.

Obsidian became the editor, and the folder became the system.

Four inboxes became one folder of plain Markdown files, with an assistant that can read and write in it.

Moving in, with help

On 26 February I told Claude, more or less: the plan is to move to Obsidian as the main organising tool — PARA, second brain. The next hours set the tone for what followed.

Claude converted 45 Word documents into Markdown, installed a connector (an MCP server, in the jargon) to read and write in the new vault itself, and planned the move of the OneNote notebooks. The next morning I asked it for a printable Obsidian cheat sheet, because I still did not know the keyboard shortcuts.

So the assistant had hands in my notes before I knew my way around them: from the first day, the vault was a workspace we shared, not a store I now and then asked a chatbot about.

A lesson from day two

I do not remember this one; my notes do. On 27 February one of the first automations, a script forwarding mail into the new setup, logged “Forwarded” for every message while nothing arrived. The log was accurate about what the script had done and silent about whether it had worked.

The lesson written down that day heads what is now a long list: working is checked where the thing is supposed to arrive, not where it was sent from. It comes back often in this series.

The pattern

Plain text as the durable layer, and an assistant with access to it. The app is replaceable — I later replaced much of what Obsidian did for me. The files stay, and they keep me independent: plain text belongs to nobody but me, and no vendor, price change or discontinued app can take it away.

Most of this series comes from that access. An assistant that can write in your notes can also write the wrong thing, in the wrong place, in the middle of the night. Deciding what it may do, and how I find out what it did, turned out to be the real work.

Part 1 of twelve in the series on how I built my AI OS. Why I am writing it: the preface.

Next: one month in, almost nothing ran where it should have — so I stopped, took stock, and let the AI interview me for my own requirements.

* Host names are changed (here: Yavin, my home server). The real ones come from a different saga.

Appendix — for builders

Start with one folder, not an app comparison. I used PARA because it already existed in three places; with the AI as the folder’s main user, any structure will do. Put everything in Markdown.
Let the assistant do the conversion. Converting Word, OneNote or PDF exports to Markdown is tedious, checkable work an AI does well; spot-check a handful of files rather than trusting its summary.
Give access deliberately. A connector (an MCP server) that reads and writes the folder turns a chatbot into a collaborator. Decide early where it may write.
Use version control from the start. I added git months later; Obsidian’s own file history works too. Either way, every change the assistant makes can be seen and undone.
Keep a dated journal page a day. One line per topic became the backbone of everything after, including this series.
Define “working” at the destination. For every pipeline (a chain of steps that moves something somewhere, such as mail into your notes), write down how you would see that it arrived.
Learn from people who show their setups. Mine were Tiago Forte (PARA) and Nick Milo (Obsidian). Follow one teacher per step and copy the structure before inventing your own.
What I considered instead: staying with OneNote (a closed format an assistant can barely reach), Notion (records, not files), or tidying the four places (they would have stayed four). I chose Markdown in one folder because the format is open, cheap for an AI to read, and mine whatever happens to any app.

Also published on Kalli Family.