The two years I wasted
I found Obsidian about two years ago. Like most people, I went to YouTube first.
What I found there was a whole world of people building note systems. Colour-coded graph views that look like a galaxy. Twenty plugins. Dataview queries. Naming schemes with three letters and a number. Whole videos about which folder structure is the correct one.
I copied a lot of it. I had a vault with tags, templates, a daily note that asked me five questions, and a graph view that looked genuinely impressive when I showed it to someone.
And it did nothing.
I still forgot todos. I still lost decisions from meetings. I still had four different places where a project lived: my head, my inbox, a Teams chat, and a note I never opened again. The system looked like it was working. It wasn’t.
Here is what I think is going on with most of that content. Building a note system is fun. Using one is not. So people build, and film the building, and the audience builds along with them, and nobody actually gets more organised. It is a hobby dressed up as productivity.
The honest test is simple. Open your vault. When did you last read a note you wrote three months ago? If the answer is never, the system is decoration.
The question that changed it
I stopped asking “how should I organise my notes?”
I started asking “what work can my notes do for me?”
That is a completely different question. Organising is passive. You file something and hope future-you finds it. Work is active. Something happens because the note exists.
Once you point a language model at a folder of plain text files, notes stop being storage. They become the memory of an assistant that can actually do things on your machine.
That is the whole idea. Obsidian is not the clever part. Obsidian is just markdown files in a folder, which is exactly why it works. No database, no export problem, no vendor. A model can read them, write them, move them and link them with no API in the middle.
What my vault does today
I run Obsidian on my MacBook with a local model wired into it. Not a chat window on the side. It has the vault, it has my instructions, and it has access to the tools on my Mac.
Here is what it does for me, day to day.
It turns raw notes into real notes
I do not format anything. During a meeting I type badly, in half sentences, sometimes in Dutch and English mixed. I dump it in one folder called Inbox.
Later I say “process my inbox”. It reads each dump and turns it into a proper note: the right filename with the date, the type and the subject, a short summary at the top, the decisions written out, and every action item as a checkbox with an owner. Then it files it in the right folder and links it to the project it belongs to.
The point is not tidiness. The point is that context travels with the note. Six weeks later I can ask what happened and get an answer, because the note knows which project it belongs to and who was in the room.
It turns minutes into something I can send
Meeting minutes are the job nobody wants. Mine used to sit as a raw transcript that I never cleaned up, which means the meeting effectively didn’t happen.
Now the raw transcript goes in and out comes a clean set of minutes: what was discussed, what was decided, what is open, who owns what and by when. Written in plain English, short enough that people actually read it.
It pulls out my todos and puts them where I’ll see them
From any meeting note it finds the things I committed to, or that someone assigned to me, and creates them as reminders on my Mac. It knows the difference between my action and someone else’s. It does not create the same one twice if I run it again.
It also keeps one rolled-up list in the vault of every open todo, grouped by project. That is the list I check when I want to know where I actually stand.
It drafts my emails
I say what I want to say, roughly. It writes the email and opens it in Outlook, ready for me to read.
It does not send. That is on purpose and I would not build it any other way. My name is on the message, so I read every one before it goes.
This one fixes a real weakness of mine. I under-communicate, especially when things are heavy. The cost of writing the email was always what stopped me. Now the cost is close to zero, so I send it.
It books my meetings
Same pattern. I say who, roughly when, and what it’s about. It opens a calendar invite in my calendar, prefilled with the subject, the people, the time and an agenda already written out. I check it and send it.
An agenda in the invite sounds small. It changes the meeting completely. People turn up knowing what we are deciding.
It answers questions about my own work
This is the part I underestimated.
“What did we decide about the firewall?”
“What am I still waiting on from Peter?”
“What do I know about the performance work?”
“Which stakeholder haven’t I spoken to in three weeks?”
I ask in normal words. It reads the notes and answers. No searching, no remembering which note it was in.
That last question is the one that earns its place. I go quiet on people under stress, exactly when I should be talking more. That has cost me before. Now something watches for it and tells me.
It creates tickets
For work items, it interviews me for the details it needs, then creates the story and its tasks in Jira directly. The interview matters more than the automation. It asks me the questions I skip when I’m in a hurry.
Why it runs locally
I am a Principal Engineer at a health company. The platform I’m responsible for serves over four million people. My notes are full of meeting content, names, architecture, security work and internal decisions.
None of that goes to someone else’s server. So the model runs on my machine, on my hardware, offline capable. The vault sits in a folder I control.
Two things follow from that, and both are worth saying plainly.
First, a local model is not as sharp as the biggest hosted ones. For this job it does not need to be. Filing a note, extracting todos from a transcript, drafting an email in my style: these are ordinary language tasks. A mid-size local model handles them well.
Second, and more important: I would rather have a slightly less clever assistant than a very clever data leak. If you work in healthcare, finance, law, or anywhere with real people’s information in your notes, this is not a preference. It is the requirement, and it should be the first decision you make, not the last.
How it is actually built
This is the part the YouTube videos would make complicated. It isn’t.
A folder of markdown files. That’s the vault. Nothing exotic.
One folder for raw input. Everything unprocessed lands in Inbox. It gets emptied. That single rule does more than any plugin I ever installed.
One naming rule. YYYY-MM-DD - Type - Subject.md. Date first so things sort themselves. Nothing else to think about.
A written instructions file. This is the real engine and the thing almost nobody talks about. It is a plain text file in the vault that tells the model who I am, what I’m responsible for, how I want things filed, how I want it to write, and what it must never do.
Mine says things like: be blunt, challenge me, don’t sugarcoat. Write in plain English, short sentences, use the everyday word. Never edit an existing note without asking. Never put real patient data in this vault. Tell me when I’ve gone quiet on someone.
That file is why the output sounds like me and not like a chatbot. If you take one thing from this article, take that. The model is not the differentiator. Your written instructions are.
Small task scripts. Each job it can do is a short set of instructions of its own: how to draft an email, how to build a calendar invite, how to extract my action items. Each one is a few lines telling the model the steps and the rules. They are all readable, editable text. When one gets something wrong, I fix the text.
Tools on the Mac. It uses what is already there. Outlook for mail. My calendar for invites. Reminders for todos. I did not build a new app. I connected the ones I already live in.
That’s the whole architecture. There is no clever bit. Anyone reading this could build it.
What doesn’t work
I’d rather tell you this now than have you find out.
It will not fix a habit you don’t have. If you don’t dump your notes somewhere, there is nothing to process. The intake habit is the whole system. Everything else is downstream of it. Mine took a few weeks to stick and I still drop it sometimes.
It gets things wrong. It mislabels a todo. It attributes a decision to the wrong person. It writes an email slightly too formal. That is why nothing it produces gets sent or committed without me looking. Draft, then human. Always.
It is not a thinking replacement. It is very good at the work around thinking: filing, retrieving, summarising, chasing. The actual hard calls are still mine. What it gives me is the state of things, so I make those calls with full information instead of half.
Setup is a day, not an evening. Not because it’s hard, but because writing down how you actually work takes longer than you expect. Most of that day is you deciding what you want, not configuring anything.
Where to start
If your vault is currently a museum, do this instead of installing anything.
- Make one folder called
Inbox. Dump every raw note there for two weeks. No formatting, no thinking. Just dump. - Write one file that explains who you are, what you work on, and how you want notes handled. Write it like you’re briefing a new assistant on day one. Be specific about your bad habits.
- Get a local model running and point it at the folder.
- Give it one job. Just one. “Turn my inbox dumps into filed notes with the todos pulled out.” Live with that for a week.
- Add the next job only when the first one is boring.
Do not start with five automations. Do not start with a folder structure. Start with the dump folder and one job.
What actually changed
I don’t take better notes than I did two years ago. My notes are worse, honestly. Messier, faster, less complete.
What changed is that nothing falls through any more. I know what I owe people. I know what we decided and why. I know who I haven’t spoken to. The emails I used to put off for a week get sent the same day.
That was always the goal. It was never the graph view.
The tools have been sitting there the whole time. Markdown files, a model you can run on your own laptop, and the apps already on your Mac. The reason it works is not the technology. It is that I finally wrote down how I work and let something else do the parts I’m bad at.