Eric Michaud runs his whole business out of plain markdown files in Obsidian, with AI agents on one side of the vault and his own ideas on the other.
Posted
4 days ago
Duration
Format
Tutorial
educational
Views
222.7K
3.2K likes
57 · 43
Big Idea
The argument in one line.
Treat an Obsidian vault as a backend database, with front matter as columns, agent.md files as maps, and a hard human/machine split, instead of a second brain you only feed.
Who This Is For
Read if. Skip if.
READ IF YOU ARE…
You keep an Obsidian or markdown vault full of notes and AI chats that you never open again, and you want it to work for you.
You run AI agents like Claude Code or Codex against your files and they keep rewriting or summarizing things you wanted left alone.
You are a solo creator who wants one place for daily metrics across YouTube, Instagram, TikTok and Skool without paying for another dashboard.
SKIP IF…
You want a step-by-step plugin tutorial; this is a conceptual overview and the app itself is behind a paid membership.
You do not use plain-text notes or AI agents at all.
TL;DR
The full version, fast.
Most second brains collect information that never gets used. Michaud turns his Obsidian vault into a backend by splitting it into a human side the agents may not edit and a machine side where they can write freely, putting an agent.md instruction file at the root of every folder, and having scheduled collector scripts write platform metrics into the front matter of each daily note. Agents log their own work into the daily note and keep a findings file they must read before the next run, so outputs improve over time. A custom front end reads the same files and lets him take actions like drafting email or making shorts.
Free for members
Chat with this breakdown — free.
Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.
Daily notes become a database when each note is a column and each front matter property is a row.
Metrics written by a scheduled API script are collected the same way every day and never pass through an AI's probabilistic summarizing.
Keep your ideas and the robot's ideas in separate folders, because AI tends to consolidate and summarize things you did not want summarized.
Let the agent copy and move files on the human side, but never write or edit them.
Dumping the whole vault into every chat drowns the agent the same way a stack of reference material drowns a person.
An agent.md file at the root of every folder works as a map, so the agent reads only the context the task needs.
Make agents log everything they did into today's daily note so you never have to journal it yourself.
A findings file the agent must read before each run turns a workflow into one that learns from every previous run.
Plain text files mean no export, no lock-in, and any agent that comes out next week can read and write them.
A dashboard that only shows you stuff is weaker than one wired to the same connections your agents use, so you can act from it.
Takeaway
Make your notes a database agents can use
WHAT TO LEARN
Structure beats volume: a split vault, a map file per folder, and machine-written metrics make a pile of notes into a system that compounds.
Separate human-written notes from agent output and forbid agents from editing the human side, so your original intent is never silently rewritten.
Put an instruction file in every folder so an agent reads a short map and then only the context that folder's task needs.
Collect metrics with scheduled API scripts into front matter, so the numbers are consistent and exist before any agent is even started.
Have agents log their own activity into the daily note, which removes journaling work and builds a history to roll up later.
Keep a findings file per workflow and require the agent to read it first, so each run starts from everything learned before.
Glossary
Terms worth knowing.
Front matter
A block of key-value properties at the top of a markdown note, which tools can read and write like fields in a database row.
agent.md
An instruction file at the root of a folder that tells an AI agent what it may touch, how that section works, and the ground rules.
Vault
The folder of plain text markdown files that Obsidian opens and treats as a workspace.
Findings file
A running log of results and lessons from past runs of a workflow that the agent reads before starting the next run.
The Script
Word for word.
Read-along
Don't just watch it. Burn it in.
See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.
17px
What you are looking at right now is the front -end operating system I use to run everything day to day. It's got what I'm doing, my numbers across every platform, my projects, my agents, the activity log, absolutely everything. It's all just running on Obsidian.
There's no database, no cloud provider, none of that. It's just plain files sitting on my desktop. When most people think Obsidian, they think, you know, note -taking app, second brain, PKM, that kind of thing.
You collect all of this information and never touch it ever again. And that's the problem with a lot of second brain setups, right? You feed it every day, all these different inputs.
forever. Obsidian can be a lot more than that. It is a crazy extensible program.
So instead of having all this information just collect dust, I want to show you how you can turn it into a backend that feeds you instead. So I'll show you how mine works because it's a lot simpler than you might think. And once you see it, it'll change how you view Obsidian.
Oh, and make sure to stick around until the end because I show you how to get this exact app for yourself. For housekeeping and to address the Obsidian critics, yes, Obsidian is just text files. And yes, the vault is just a folder.
If I click on Obsidian, there's a file explorer right there and that's it. just because that's really all it is. Which is honestly a really good thing because I own all of it.
Nobody takes any of it away and anything I use can read and write to these plain text files, whether it's Cloud Code, Codex, Pi, Hermes, or something new that comes out next week. I don't have to ever export my memories or my chat history. Nothing's locked behind a subscription.
It's all mine and it's all in one spot. The way I like to use it is kind of a hybrid of traditional note -taking apps like Evernote or whatever it was you were using and my AI agents working in the same place. With that, I think it's incredibly important to keep...
your ideas and the robot ideas totally separate because AI has a really bad habit of consolidating and summarizing things that you might not want summarized. And you never really know if it's capturing your intent or just what it thinks your intent is. So I draw a hard line where my vault is split into human and machine.
Okay. So the human side is my side, my ideas, my context, exactly as I intended them to be. So daily notes, tasks, projects, people, my context, the agent can move things around and copy, but it's not allowed to write or edit.
anything on the human side of the vault machine is obviously the robot side and it can go nuts it's got all the permissions it needs it can write read edit i don't really care it can make a giant mess on the machine folder so this is really where all its outputs go so scripts logs research workflows that lives in there and it just does its own thing and i don't really mind too too much at the root of every folder in the vault including the vault itself there is an agent .md file and it's really just the instruction manual for that section okay what it can and can't touch how to use this particular section just getting the agent caught up on exactly like what's going on in this section here the ground rules here's how we play in this sandbox and there's a couple rules that make the entire thing work for me the entire mechanic is based around the markdown files that obsidian uses naturally and the biggest single instruction that my agents have is to log everything they do into today's daily note so that the agent is the one filling in what it did and i'm not like journaling and going back over everything every single time it fills in itself but if you look up here at the top at the properties this is a markdown front matter and we've done something similar to the auto logging with
the metrics that get filled in here. So every platform I use has its own little collector script, right? So YouTube comes in through its official API, school, Instagram, TikTok, all my social medias, all my inputs essentially, right?
They run on a schedule and then just fill in their different values into these front matters, which makes things a lot easier because I don't have to go and log this all myself. It just pulls in through the API call. And it also means that the exact same data is being collected the exact same way every single day.
And it doesn't go through AI's like probable holistic filters. So the nice thing is I can open up my daily note at any point in time and the metrics are already there.
I've already got a snapshot of absolutely everything I want to keep track of that's publicly available at least. And I don't have to worry about my AI agents getting like up to date first. Okay.
So if you think of every one of those daily notes as a column and every one of those front matter properties as a row, Obsidian stops looking like a silly note -taking app and turns into a full -on backend database. And the cool part about adding AI to this as well as every one of those automatically logged activities. can also roll up and do the exact same thing over time.
And that is a payoff. The point of this entire video is that once your data looks like that, you can build anything on top of it. And that's what's really exciting.
So I'll give you a couple of examples because I've got two things on top of this that work for me practically. First thing I want to talk about is the AI agents and how they work inside the structure, how they roll up all these different historical notes and turn it into something super valuable. The mistake a lot of people make with their agents inside these different second brain vault structures is they dump the entire context into to every chat, right?
Whether it's intentional or not, maybe they think that like more context means a better answer. What it actually does is it just drowns the agent. It's exactly the same as handing somebody a stack of reference materials, right?
You're going to get like the beginning and the end parts maybe, but all the middle bits just get kind of lost. It's too much information. Agents are exactly the same.
That's why the agents file at the top of every folder is so important. Like your AI will go in and read the vault map, essentially like the TLDR and understand exactly how to use each folder and it doesn't have to look. through everything.
Then when it gets to that folder, it'll read exactly the context it needs for the task from the agents file. So it'll pull in the workflow, the daily notes, it'll get the context it needs and know what it needs to actually achieve that goal because you've got a map there for it. Like take my little animated reels, for example, right?
That folder has all of the assets and instructions it needs to make one. But what makes it really cool is that every time we make one, we write notes, performance numbers, which topics hit, which gags worked, like that's non -negotiable. It tracks what we've done already.
That all goes into a findings file that the agent has to learn from before the next run. So it's not just like, here's a workflow to make a short. It's here's the workflow.
Here's where we're at. And here's everything we learned from doing this a bajillion times in the past. So every short gets better than the last one.
This structure alone has been like super helpful, right? Like my agents are always getting better. It's given me clarity and perspective on what I should be doing.
And if you stopped here, that's fine. Like you could just run this in Obsidian or in Cloud Code or whatever with just the terminal. I spent a long time just using the terminal like that as well, basically just like cloud code or my PI harness in there.
But I kind of wanted something cooler and something easier to work out of. And this is the front end that you saw at the top. I took a lot of cues from the original Apple Music app.
I don't know if you guys remember, but like you were like scrolling through. album covers and stuff or like ps5 steam os really ux forward applications and i put it into my own thing i'm like more inclined to use something if it looks cool this is doing all the same things we were doing in the terminal like it's reading the same markdown files it's running the same scripts it's got the same connections this is just cool because you can arrange it the way that you want it to it's really great for ergonomics so this changes through the day it pulls in what i've got to do has my email available and shows me like what needs attention and because it's got all that history it knows my highest leverage activities and can make educating suggestions as to like what comes next, which I put into a dropdown called needs you.
And these are the things that actually move the needle that actually need my attention based on the history sitting in these properties. What sets this apart from what I see a lot of other people doing is this isn't just a dashboard that shows you stuff. You can actually take action through here as well, because we've got all the same connections that I was talking about before.
So needs you will say something like maybe I need to draft an email. I've got Gmail tied in so I can actually draft and send that through this one dashboard or whatever action. I need to take as long as I've connected that on the backend.
All of my different workflows I can do through here. I can create shorts through here. I can do my content research through here.
You can fatten this out to do anything your AI agents were doing for you already. For example, my content tab, I can grab like five or 10 different videos and see what people are talking about. And then just like sick my AI agents on that topic.
It'll scrape the comments and just get a feel for what people are saying. And then I can narrow in on the topics that like suit me. It can make suggestions about what works for me too.
And it's not just making it up because it's got historical. data about what works on my channel. If this all looks complicated, it's really not.
It's just a lot of small things kind of stacked on top of each other. But the thousand foot overview is markdown files with front matter properties, making sure to have an agent's file at the top of every folder, and then making sure that you have some way of keeping track of findings and you can improve your agents over time.
From there, there's really nothing you can't build on top of this. And a lot of folks are very, very close using Obsidian as like a visualizer, second brain. But if you take a step back and just picture it as a backend, it opens up so many more possibilities for you.
If you want to learn to build like this for yourself, I've put together a lot of resources in my school community where you can go in and learn how to build your own Obsidian plugins. Also, I've put together the exact app that I was showing you in this video. It's available through the VIP membership.
If you want to check it out first, there's a clickable demo. I'll leave a link to that in the description as well. Thank you so much for hanging out with me.
Make sure to check out this video and I'll see you in the next one.
The Hook
The bait, then the rug-pull.
Everything on screen runs from plain files on a desktop, with no database and no cloud provider. Michaud's pitch is that Obsidian is a backend, and the second brain framing is why most vaults go unused.
Add Modern Creator as a preferred source and Google shows you more of our breakdowns in Search, Top Stories, and AI Overviews. It only changes what you see, and you can undo it in your Google settings anytime.
Add to Preferred SourcesOpens your Google source preferences with us pre-loaded. Tick the box and you're done.
A 7-minute live-vault demo arguing that Obsidian is not a note-taking app but a unified AI operating system where capture, execution, and intelligence compound under one roof.
00:00
00:09
00:15
00:22
00:28
00:34
00:41
00:45
00:53
01:00
01:06
01:13
01:19
01:25
01:31
01:37
01:44
01:50
01:56
02:03
02:09
02:15
02:22
02:28
02:34
02:40
02:47
02:53
02:59
03:06
03:12
03:18
03:25
03:31
03:37
03:44
03:50
03:57
04:03
04:09
04:17
04:21
04:28
04:34
04:40
04:47
04:53
04:59
05:06
05:12
05:18
05:25
05:31
05:37
05:44
05:50
05:56
06:02
06:09
06:15
06:21
06:28
06:34
06:40
06:47
06:53
06:59
07:06
07:12
07:18
07:25
07:31
07:37
07:43
07:50
07:56
08:02
08:09
08:15
08:21
07:20
16:42
13:59
14:35
20:04
20:54