Nick Milo diagnoses Tiago Forte's real, in-use Obsidian vault like an X-ray, then the two of them build the small system of files and folders that makes any AI model safe to point at it.
Posted
2 weeks ago
Duration
Format
Interview
educational
Views
48.3K
1.1K likes
57 · 43
Big Idea
The argument in one line.
A note vault stops being just your second brain and becomes the AI's second brain too, once it's wrapped in a backup habit, a scoped folder boundary, and three small files that translate your working style to any model.
Who This Is For
Read if. Skip if.
READ IF YOU ARE…
You keep notes in Obsidian, Logseq, or any local markdown vault and want to connect an AI model to them without losing control of the originals.
You've started pointing ChatGPT, Claude, or a Cowork/Projects session at a folder of your own notes and want a safer, more repeatable setup.
You're deciding whether your notes need more folders, links, or structure now that AI can search unorganized text directly.
You've had AI generate a note or a map of content for you and felt buried by the output rather than helped by it.
You're curious what a real, non-demo knowledge base looks like when an outside expert reads it cold.
SKIP IF…
You don't take notes in a local file format and have no interest in file-based tools like Obsidian.
You're looking for a step-by-step Obsidian tutorial for total beginners — this assumes you already know what the app is.
TL;DR
The full version, fast.
Tiago Forte lets Nick Milo read his actual book-writing Obsidian vault cold, like a doctor scanning an X-ray. From there they build a practical system for letting AI work inside a note vault: always back up before granting access, since read access implies write and delete risk; point AI at the smallest folder that has the context it needs rather than the whole vault; and use three small text files, a 'me' file, a vault map, and a skill map, as a translation layer any AI model can read. Folders and links still matter because they signal to AI which notes carry the most 'gravity,' the same way PageRank inferred importance from links. A single AI-only folder (their AIOS/history folder) keeps every AI-generated file quarantined from your own thinking, and a personal rule against generating more text than you can actually read prevents AI from burying you in cognitive debt.
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.
Tiago hands Nick his real vault cold and Nick reads it like a diagnosis: theme, structure, and how vanilla and focused it looks.
01:30 – 02:47
02 · What Obsidian really is
A quick refresher: Obsidian holds nothing itself, it's a skin over one folder of plain markdown text files that any app or AI can be pointed at.
02:47 – 07:15
03 · Backup first
Tiago's rule is AI never touches the originals; Nick's is more flexible but rests on the same principle, that read access implies write and delete risk, so back up before granting any access.
07:15 – 10:50
04 · What's the safest way to connect AI to the vault?
Of the three ways to connect AI (point it at a folder, a third-party plugin, or the Obsidian CLI), pointing a Cowork or Projects session at a folder wins for having the fewest moving parts.
10:50 – 11:27
05 · The folders are back, baby
A folder stops being just a human filing tool and becomes a literal boundary defining what an AI model is allowed to see and touch.
11:27 – 12:22
06 · Nick's 3 files model
Nick's whole system for onboarding an AI model to his notes is three files: a 'me' file, a vault map file, and a skill map.
12:22 – 16:37
07 · Own the land you think on
The outer AI layer is the least stable part of the system and shouldn't be the center of the workflow; your notes are the core, AI models are interchangeable plug-ins, which is digital sovereignty over digital feudalism.
16:37 – 19:26
08 · Why AI still needs your folders and links
Past a certain vault size, models sample rather than read everything, and links act like PageRank, telling the model which notes carry the most importance; Nick's rule of thumb for that threshold is 500 notes.
19:26 – 21:50
09 · 2 tricks worth stealing
Hand AI the exact system-root file path instead of describing a note, and instead of describing your taste abstractly, feed AI your 15-20 best documents and have it reverse-engineer a style guide (the Sherpa system).
21:50 – 23:25
10 · The cognitive debt
Every wall of AI-generated text is a decision a human still has to process, and unread output piles up like debt against your future attention.
23:25 – 25:52
11 · Organizing your notes is a way of thinking, not filing
Mapping ideas by hand is itself a meaning-making process, not just an artifact to produce, and that's a reason to keep organizing even when AI can search a messy pile of notes just fine.
25:52 – 29:01
12 · One folder for everything AI creates
A single dedicated folder (their AIOS/history folder) is the only place AI is allowed to write new notes, keeping a hard line between a person's own thinking and anything AI has touched.
29:01 – 32:50
13 · Put AI's wikilinks to a real test
They task AI with linking a new project dossier to 84 existing notes and building a map of content section, then watch it work through the vault live.
32:50 – 35:03
14 · The pocket-universe problem
Given a short, incomplete list of email topics, AI treats it as the entire truth and over-indexes on it, because a model only knows the closed 'pocket universe' of context it's been handed.
35:03 – 37:27
15 · File over app: your vault is the AI's second brain now
Because AI has no persistent memory of its own, a note vault now has to serve as both the human's second brain and the AI's, making portable, file-based notes more valuable than ever.
Atomic Insights
Lines worth screenshotting.
If AI can read a file, you have to assume it can eventually write to and delete it too, even without explicit write permission, so back up before granting any access.
The safest default for connecting AI to notes isn't a plugin or a CLI, it's just pointing a Cowork or Projects session at a folder.
Give AI the smallest subfolder that has the context it needs for the task, not your entire vault, when the material is sensitive.
A folder is no longer just a human filing system, it's a boundary that defines exactly what an AI model is allowed to see and touch.
The translation layer between a person's notes and any AI model can be just three text files: a 'me' file, a vault map, and a skill map.
Betting your workflow on one AI model's outer-layer features is a mistake, because the model is the most ephemeral part of the stack and your notes are the part worth keeping stable.
A note with 50-plus incoming links is a signal to an AI model that the note carries more importance, the same way PageRank inferred a webpage's importance from links pointing to it.
Past roughly 500 notes, a model stops reading an entire vault and starts sampling it, and folders, links, and metadata are what tell it how to sample.
When you want AI to work from one exact note, hand it the file's absolute system-root path instead of describing the note, so no tokens or time are wasted on ambiguity.
Instead of describing your taste or style abstractly, hand AI 15 to 20 examples of your best work and have it reverse-engineer the pattern into a reusable style guide.
Every AI-generated note or summary creates cognitive debt, since a human still has to read, evaluate, and decide what to do with it, and unread AI output is a loan against future attention.
A single dedicated folder for everything AI creates keeps a hard line between a person's own thinking and anything a model has touched, so the vault never turns into machine-generated slop.
AI treats whatever context it's given as the entire truth, so a short or incomplete list handed to a model gets over-indexed on as if nothing else exists.
Organizing notes by hand isn't obsolete just because AI can search unorganized text; the act of organizing is itself a way of thinking, not merely a filing task.
A note vault used to be only a human's second brain, but because AI has no persistent memory of its own, the same vault now has to serve as the AI's second brain too.
Takeaway
Three ways to make an AI-touched note vault safe and useful
WHAT TO LEARN
Before letting any AI model near your real notes, back it up, scope what it can see to a folder, and give it a small set of files that translate your working style so it doesn't have to guess.
01Nick Milo reads my vault
A vault built around one active project with almost no distractions signals real focus, which is itself useful diagnostic information about how someone works.
03Backup first
If AI can read a file, assume it can eventually write to or delete it too, so back up your notes before granting any access, whether manually, automatically, or off-device.
Spreading backups across more than one location and off the original device lowers the real-world risk of pointing AI at your notes.
04What's the safest way to connect AI to the vault?
Point AI at the smallest folder or subfolder that has the context it needs for the task, rather than your entire vault, especially when the material is sensitive.
The simplest, lowest-maintenance way to connect AI to your notes is pointing a Cowork or Projects session directly at a folder, rather than relying on third-party plugins or a CLI.
06Nick's 3 files model
Build a small translation layer of a few text files, who you are, how your notes are structured, and what tasks or templates to use, so any AI model can pick up your working style without re-explaining it every session.
07Own the land you think on
Treat the specific AI model you're using as the least important, most replaceable part of your system; your notes are the core asset, and models should be swappable without disrupting your work.
08Why AI still needs your folders and links
A note's number of incoming links functions like PageRank, signaling to an AI model which notes carry more importance once a vault gets too large for the model to read in full.
Once a vault is large, roughly 500 notes as a rule of thumb, links, folders, and metadata become the signal that tells AI what to sample, not just a nice-to-have for humans.
092 tricks worth stealing
Hand AI the exact system-root file path of a note instead of describing it, to guarantee it starts from the right file without wasting tokens or time.
To teach AI your taste or style, give it 15 to 20 of your best existing examples and have it reverse-engineer the pattern, rather than trying to describe it abstractly.
10The cognitive debt
Every piece of AI-generated text creates cognitive debt, a decision a human still has to read and process, so don't generate more than you actually have the time or energy to read.
11Organizing your notes is a way of thinking, not filing
Organizing notes by hand, through mapping and linking, is itself a form of thinking and meaning-making, not just filing, which is a reason to keep doing it even when AI can search unorganized notes just fine.
12One folder for everything AI creates
Keep one dedicated folder as the only place AI is allowed to write new notes, so everything AI has touched stays quarantined and separable from your own thinking.
14The pocket-universe problem
AI treats whatever context it's given as the complete truth and will over-index on a short or incomplete list, so check what information it wasn't given before trusting its output.
15File over app: your vault is the AI's second brain now
Because AI has no persistent memory of its own, a personal note vault now has to function as both your second brain and the AI's, which makes portable, file-based notes more valuable, not less.
Glossary
Terms worth knowing.
Vault
Obsidian's term for a folder of plain markdown text files on your own computer that the app displays and links, rather than storing or owning itself.
File over app
The principle that your notes should exist as plain files you own and can move anywhere, rather than being locked inside one company's proprietary app or format.
Cowork
A mode inside an AI chat tool (used here in an AI assistant) that lets the model actively work inside a pointed-to project or folder, rather than just answering one-off chat messages.
AIOS / translation layer
Nick Milo's term for the small set of files (a 'me' file, a vault map, and a skill map) that tell any AI model who you are, how your notes are structured, and what skills or templates to use.
Map of Content (MOC)
A single note that gathers and links related ideas together, acting as an index or overview page for a topic inside a note vault.
Cognitive debt
The buildup of unread or undecided AI-generated output that a person still has to process later, functioning like a loan taken out against future attention.
Pocket universe problem
The tendency of an AI model to treat whatever context or files it's been given in a session as the complete and total reality, rather than a partial slice of a larger truth.
Sherpa system
A method where you give AI your best 15 to 20 existing documents so it can reverse-engineer a style guide from them, instead of trying to describe your taste abstractly.
System root path
The absolute file path to a specific document on your computer, used to point AI at an exact file with no ambiguity about which note is meant.
Resources
Things they pointed at.
17:12toolCopilot (Obsidian community plugin)
17:12toolSmart Connections (Obsidian community plugin)
Quotables
Lines you could clip.
00:21
“When you see a vault open up, what do you notice? Like almost like a doctor, like what's your diagnosis?”
the framing device for the entire video, works as a cold open for any note-taking or PKM short→ TikTok hook↗ Tweet quote
10:36
“If AI can read your entire vault in one go, what is a folder even for?”
sharp rhetorical question that reframes the whole video's thesis→ IG reel cold open↗ Tweet quote
“How can we actually own the land that we think on, as opposed to digital feudalism when we're doing all of our work on somebody else's land or in somebody else's software?”
the video's core metaphor, land ownership vs. renting, in one self-contained line→ newsletter pull-quote↗ Tweet quote
22:48
“You're piling up cognitive debt. You're taking out a loan against your future attention.”
a named, memorable concept that stands alone as advice about AI-generated content→ IG reel cold open↗ Tweet quote
26:15
“How do you stop it from turning your vault into a landfill of machine-generated slop?”
vivid, relatable problem statement for anyone using AI to write notes→ TikTok hook↗ Tweet quote
35:04
“Your vault isn't just your second brain anymore. It's also the AI's second brain.”
the closing thesis line, works as a title card or pull-quote→ newsletter pull-quote↗ Tweet quote
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
metaphoranalogy
Every time I post about Obsidian, the same kinds of questions come back. I've been collecting these questions for months. So I brought them to the person who's been living at this intersection of Obsidian and AI for longer than anyone I know.
I sat him down in front of my vault. When you see a vault open up, what do you notice? like almost like a doctor like what's what is your diagnosis first thing i'm thinking is what theme is it so it's like default dark theme so you already know i'm basic yes if you spent any time around note -taking online you probably already know nick milo he's been on this channel many times before and he is the person i go to when i want to think through this properly People want to use AI, obviously.
They want to own and control and know where their data is. That's Obsidian, file over app. And you have been living and working at that intersection longer than anyone.
And I feel like I've thought more deeply than anyone I know. That is why he got my real life vault, not just a demo. Everything in my Obsidian vault revolves around my new book.
So it's a pretty vanilla. setup, like very basic, which is great. You're not getting bogged down in the million distractions that Obsidian allows you to be distracted by.
So to me, this tells me that you're really focused on your work at hand. Let's do just a minute of housekeeping for anyone who doesn't already use Obsidian, because the rest of this video assumes that you kind of know how it works. What is Obsidian?
And at its basic form, it's a note app. There are many note apps. I mean, I could go like philosophical, but it's just, it's a note app.
The difference I think is like with Evernote, Evernote holds the notes. Notion holds the notes. Obsidian is actually not holding anything.
It's only like a skin that you put on because your computer is holding the notes. It's one folder with four subfolders, clippings, Copilot, Readwise, Web Clipper. That's it.
And then within those folders are text files. So super basic, that's what people who use Obsidian really appreciate about it because they don't have to be locked into Obsidian. That's why you can point anything at them, any LLM, any AI, including ones that haven't even been invented yet.
So we found that Nick and I have two separate rules that are similar but also different. My rule is that AI never touches the originals. If I want to work with an LLM, I duplicate the vault and I point the LLM at the copy.
His turns out to be a little more nuanced and subtle and I would say flexible. This graph you see is the result of Not as many hours as you spent, but hundreds, hundreds and hundreds of hours.
All the linking, the connecting, there's tons of manual notes, there's tons of outlines with structure was all my personal time. Do you recommend people point AI at the canonical notes or create a duplicate or what? If I were to point AI at this safely, there are a few different ways to do it.
The first one, always make a backup. Because if AI can... read something, you have to assume, even if you don't have the permission for it to write, you have to assume it can write, which means it can delete.
It can change things. Have you seen instances of that where it technically had read only and something was edited? I've heard of those instances.
I mean, it's not affected me. And it just depends on your tolerance for risk. So I have pointed my AI at all of my notes, but it's because I have a backup system.
So you have to back up. I do manual. Will do automatic backups and then off of the same computer as well like hey just throw one into Google Drive, that's a zip file put one on a zip drive You know, you're just spreading out your risk that way so I haven't been worried about it like maliciously editing things, but I've heard cases where that's happened.
So whereas I use duplicating the vault as a first step, he uses it as a kind of last resort. But I would say these share the same underlying principle. Before giving the AI access to anything, back it up.
When I talk about my AI mirror vault, i'm not referring to a autonomously maintained self -syncing system with a complex diagram workflow it's really just a copy of a folder that's the whole thing it takes two seconds and i want to show you that kind of boring approach that boring version because that's where you should start as well so here i have my finder i have my resources and there you see my obsidian vault cool so let's assume that i want to create a duplicate Is this really just as simple as right -click duplicate?
I would just hit command D. Even faster. Which I know that you know.
Yes. You're good with hotkeys. It almost seems like that's too simple.
It's hard to believe that it's that easy. And then I would just do 07, 09, am I doing that right? Or 19.
Is that something you typically do is date it? Yeah. And I'll just do that.
And then I end up having... you know a backups folder and that's just where i can kind of loosely do backups of the on -device variety but again we want to also have one that's not just on the same device okay but yeah it's as simple as that and if we open this up for anyone who's not familiar so this was a complete perfect duplicated copy of my entire vault right so these are markdown files which are just plain text not encrypted no special proprietary format it's just plain text and that's like one of the big selling points of using an app like obsidian and let's really look at this so if we go back we just made a duplicate like how many duplicates until i fill up my computer and if we actually look at the size of this that's what are we 120 megabytes a little bit larger than i was thinking so you probably might have some media maybe something with readwise is pulling in maybe i mean if they're just the plain text variety then we're talking like 12 megabytes yeah exactly a tenth of i probably have photos some photos in there that makes sense even so that's pretty small that's tiny yeah it's nice that i can do it so easily because then you can almost create throwaway duplicates like if you want to try some crazy experiment yes you just create another one typically what i'll do so it'll be like um pre nick session like i'll just do things like that and that's good
That's good. Okay. So then we don't have to be precious about this one.
We don't have to be scared. We can go crazy. It's just command D, date the copy and you're done.
Although this approach is simple, the copy you create does not stay updated. It's a single point in time snapshot. So it starts going stale the moment you take it.
And the only way to update it is by making a new one. There are three ways to connect AI to your notes. i had to ask him which one he uses there's so many ways to connect ai to your obsidian notes now the three that i know of are just point say cloud co -work at the at the directory at the folder but then there's an obsidian cli now and then there's like a whole variety of plugins that integrate ai yeah how do you think what is the what is the taxonomy of these different or the preference that you have for them just point in ai at the folder yeah i mean i'll talk through the other options but at the end of the day if you use a third -party plugin and there are a couple great ones there's a third -party plugin in obsidian you can download or community plugin it's called copilot that's very cool there's another one that's a little bit different called smart connections sure use those but now you're you're dealing with a few developers and they're awesome but they're separate developers so how are they going to maintain that
all this extra, all these extra breaking points. So it's like, what result are you after? You want AI to be able to look at your notes.
You want to be able to work with it. The simplest thing to do is point something like a cloud co -work or GPT work and just point it at this, you know, whatever folder these two that you'd prefer to use. Now, there are ways you can like have the third party.
plugins and then you can ask ai questions from from inside of obsidian i just i don't want these multiple ways to do things that exists and then you know we obsidian cli i don't even want to touch that i know it's extremely powerful but until i have time and a really a really good reason to do like the coding that's involved to figure it out i don't need it He isn't against plugins, he's against depending on them as a core part of your system.
When you have fewer moving parts, fewer critical pieces, that means fewer things that can break as soon as some random person stops maintaining the repository. The most common question I get about all of this, how do I do this without shipping my second brain off to a data center somewhere? My straight answer is we did not do any of these demos locally.
A cloud -based model was reading my vault the whole time. But there is a way that you can control what the LLM has access to. So I'm just going to open up a new co -work session, make sure it's on co -work, which is much more powerful.
So I just click here. And that's the magic button, really. It's just project or folder, right?
Yeah. And that's the control mechanism. So when we talk about privacy and owning your notes, because you don't have to point it even at your main Obsidian vault.
You could say, only the notes that I've manually put into the subfolder. And then you point your AI at the subfolder. So for those that are a little bit more...
private with what they're sharing they're like no only the subfolder with these 10 notes and it's like you have that granular control at any moment with how much you want to share or not it's about limiting the scope that it can touch not about trying to encrypt everything the ai sees only the folder or the subfolder that you hand it that you give it permission to access and it can't see or touch or edit anything outside of it So a principle here is to give it not necessarily your entire obsidian vault if privacy and security is an issue, but to give it the smallest subfolder that has the context it needs for the task at hand.
If the material is sensitive, that's how I'd start. If AI can read your entire vault in one go, what is a folder even for? So this is super interesting because this implies that you need folders.
Yes. Is there any way to do what you just described with anything other than folders? Folders are back, baby.
Folders are back. You can think of a folder in the modern age as just a boundary, a container for what the AI is allowed to see and touch. It's not necessarily a requirement for how you think as a human, but what was once a container for human thought is now also a container for the AI's intelligence.
Nick has a model for how these pieces fit together. And it's really made up of just three pretty simple parts. Whenever I start a new chat co -work, doesn't matter.
It's, hey, look at my me file. It tells it like, here's who I am. Look at my vault map file.
Here's how the folder structures are. Here's how you create new notes. Like if I want it to be a project note versus an idea note or something, it knows the templates.
So it can do all that for me. I just say, hey, create a new project on this. and then the third one is the skill map the skill map is um you know we can use skills with ai and these are like defined ones that then get to the next question it's like what can we do and so what happens is there's it's it's like you say it's the just in time situation and it's like okay what's the problem that i'm now really feeling urgent about or that's coming up let's say let's say a little bit more about this yeah This is what you call your AIOS.
Like call it whatever you want, but it's the idea that there are three layers to how we can interact with AI. So at the core of it is a second brain, a knowledge system, an idea verse, your notes. That's at the core.
Everyone has that. Before AI was around, those who had notes, we had this. AI is around.
So now what do we do? Well, there's the outer layer. go skip the middle layer there's the outer layer that is clod that's open ai chat gpt you know what else do we have like any of these models that are out there gemini frank you have to launch a a product called frank now you plugged it so much i know that's funny it's like my go -to made -up ai model frank And so the outer layer is the least important.
It's the most ephemeral. It's always changing. One month it's Claude Cowork and then something else will come along and we'll go, ooh, I want to play with that.
But I'm stuck in Cowork. Well, let's not get stuck anymore. And that's one of the value propositions of having your notes on your computer.
Like the Notion apps, the other apps, they are great and they're built in. So it's a little bit easier to use AI features that way. But the advantage of something where you can use Obsidian to look at the notes on your computer.
is that you can use any AI at any time. They increase their prices. Well, that's okay.
You are not actually locked in. I haven't thought of it this way before, but I feel like you're inverting the whole model. The way most people, like really almost everyone thinks is, okay, which AI should I use?
And that's at the center, even if they don't realize it. And then they think, oh, let me add some integrations or connectors or a working directory. But just that way of thinking and speaking, your data is sort of at the periphery.
It's like, oh, it's plugging in and it's enriching the answers. But no, you're inverting that and you're saying your knowledge and information and data is the core. And the AI is almost like these little robots that just like come in and just like plug in and interface at different points, but they're not essential to the system.
And doesn't that seem like how it should be? Yeah. And the reason it's not is because the tech companies, rightly so, That doesn't give them the same revenue growth, knowing that you could leave their platform at any time.
So just incentive -based, their incentive is to lock you in if they can do that. So the idea behind this, and we'll get back to the three layers, but it's digital sovereignty. How can we actually own the land that we think on as opposed to digital feudalism when we're doing all of our work on somebody else's land or in somebody else's software?
So in the middle, your knowledge, the outside layer is the tech layer, the AI. So we need some way to connect the two to translate. And it's not complex at all.
That's what I mean, just three files. Those three that you mentioned. Yes, that is the translation layer.
It's just a playbook. It's an operating manual. It's, hey, AI, it doesn't matter who you are.
Here's how you're going to work with me, the individual. It's that, I mean, so we can talk big about it, but at the end of the day, it's just, you have a knowledge here. Choose any tool you want and then read these files.
Now we can work together. I love the way that Nick puts this. He says, own the land you think of.
That's such a wonderful way to think about what you're trying to accomplish here. And it doesn't require a ton of work. What Nick calls the translation layer that sits between your notes and the AI is just three text files.
Totally doable. At first, my reaction to this three file translation layer was kind of, why bother? is this really necessary the models are so good already and getting better all the time can't they come up with those three files can't they make the translation layer themselves can't i just point it at the vault in its entirety and ask it to make the translation layer i know that my own master prompt has a lot of personal context and professional context in it and it has nothing about how i structure my vault or my work and i can't think of a time that that's obviously hurt me So the closest concept that I think we have to that translation layer is the master prompt, which I do have here.
Here in my settings, there's this pretty extensive master prompt that has personal context about me. But you know what I noticed in your explanation is what's not in there is what you said about like, how do projects start? What are the structures that we turn to?
What are the best practices for how we work with information? I don't have anything. I just leave that up to the AI.
So aren't the LLMs smart enough to already know how to structure things? Now, you've done a one -off project. It was a huge one, writing a book.
The more that you've done your own notes and made links and connections, the more these links and connections are telling any AI tool, what are the important notes so that one that had 50 plus links to it it knows that is has high gravity and it can interpret that intelligently and say it has the most links so i'm going to make sure i read this note because the bigger the vaults get it's not able to read every single word so it can make assumptions assumptions are what lead to the errors or hallucinations or overconfidence and that sort of thing and we might not always know that it's not reading every bit of content.
So that's where when we do personal organization, we're telling the AI models what is meaningful, and it can read that by the amount of links that a specific note might have. Interesting. So it prioritizes the important notes.
See, I like that. It's so similar to how Google approached the internet with PageRank. Yes.
It's like instead of having thousands, millions of humans... over there cataloging let's infer meaning from the graph from the connections and now our personal note collections are large enough to need that same approach yes and not every like use case is going to need all that like complexity like it can be a lot simpler and you don't have to organize you can have zero folders too and just it depends on what the prompt is like what problem that is trying to be solved But Nick points to something that I've also noticed as well, which is past a certain size, the model is not reading your entire vault, it's sampling it.
And the structure that you give it, whether that's through folders or links or metadata, is a signal telling it what to sample and how to sample. His rule of thumb for what that threshold is, is 500 notes. We never actually tested it, but I think that's a great rule.
It's probably approximately right. for when you should start thinking about adding folders or links or metadata to your Obsidian notes. There are two things I learned from Nick during this shoot, and I used one of them on camera 30 minutes later in front of him.
The first is small but powerful. When you want the AI to refer to a specific note, don't try to just describe the note generally. Hand it the exact file path.
And there are a couple... hot keys but we can just hit command or control p and i'm going to type copy which gives me the commands that use copy system root is the most guaranteed it's ugly but it goes all the way to like your whole computer system so these are all hot keyed it's like command shift c or something for me but then like if we did the system root we can see what that looks like sometimes if i want absolute certainty i don't want to waste tokens i don't want to waste my own time even wondering if it's going to be in the exact right place, I'll just use the system root path.
And it's like, start with this note, help build a brief around this. And the second one isn't really specific to just Obsidian. When trying to describe his taste, his style, his point of view, he didn't try to describe it abstractly or in general terms.
He took a much better approach. Yeah, there's a couple skills that... Put together we called the Sherpa system and essentially I say hey train on my these are my favorite 15 to 20 maps of content i've ever made some of them are more like an index others are like a thinking space and i like you know arranged everything others are like all pros with all these links built in so it can be a lot of different varieties i want you to look at all my favorite ones and then reverse engineer what makes a good map of content and then i worked with the model and i was like now we have a style guide for maps of content interesting yeah and that's part of the sherpa system yes so that's one of the skills like think skill here and then the the running system and then so there's that skill and then the other one is like the questionnaire so so what i do is i'll just go to the message here and i'll say run the sherpa and then it taps into the system and then it starts the questionnaire skill
It's like, what topic are we learning about? This is such a powerful way to work with AI. Instead of trying to articulate what is very difficult to express, your tacit knowledge, just give the LLM examples and have it decide and summarize what makes them good.
I think one of the most underappreciated aspects of working with AI, which is something that Nick notices and mentions as well, is that it increases your cognitive debt when ai does a map of content for you it's only potential meaning and that can be disconnective if you don't connect it afterwards you know what i mean like it's just this wave of new cognitive debt it's it's not helpful until you actually have the energy to make it your own AI is constantly generating these walls of text that you have to read and understand and take action on.
It's always giving you multiple decisions, multiple paths that you could walk down. It's always giving you, oh, do you want me to do this one other thing or take this one other action? Each and every one of those decisions and choices takes up cognitive energy.
So over time, as you work with AI, you're kind of piling up cognitive debt. You're taking out a loan against your future attention.
Like any kind of debt, that's fine until the amount that you've borrowed starts to exceed what you can service, what you can pay back. So you need a rule to kind of limit that cognitive debt. The one that I've been using is don't generate anything that you don't have the time or the energy to read.
What is the point of a lot of text that is more than your human mind can absorb? Here's the question that everybody's really asking about digital notes these days. If AI can search a pile of unorganized text, of unorganized notes, then what's the point of organizing?
This is really the big question. We can move text around really fast now, and we can link to all these other things just by one thing. And we can basically work closer to the speed of thought in a single note, a single dedicated space.
that can hold all of these disparate thoughts and it's like wow finally i can just be in this space and not worry about like what am i forgetting to think about i've actually collected everything into a single note and i start to move things around quickly so i'm i'm the cartographer of my ideas of of my knowledge and so it's not so much i want to create an artifact but it's more, let me think about this topic.
How are things related? Because working in this mapping way, it's not organization for organization's sake. Organization in this way is ideation.
All organization is created equal, but I feel like through this sort of like mapping method that the organization. is a form of ideation that i love because i kind of skew towards organizing like more that architect but then in this mapping method it's gardner and i'm just feeling like like i'm the avatar of like pulling on fire and air and and water and earth together that's how i feel like in this so it's like all parts of my mind activate um If someone from the outside looks at it, they might go, oh, it kind of looks like an index.
But they can't see the meaning that you made through the making of the map. It's like you're mapping your own relationship to the topic, which is not a fixed, it's not an objective thing. It's inherently subjective.
priorities you hold the interpretations you impart there's like a pieces of you in the map almost it's a it's a meaning making process um it can be more utilitarian but the one the stuff that usually feels that makes me come alive is when i'm like whoa this is a really meaningful thing it might just be about trying to learn the guitar but if i'm mapping it myself that's a lot different than you know reading a wikipedia article it's it's um it's personal totally and i would say if organizing helps you discover what you think or what's important then it's still worth doing ai can't do that for you but i found nick's point of view striking which is that you really have to question whether organizing is needed how much organizing is needed there may be less needed now than before this is the guy whose whole business is teaching this method that's why i believe him Here's a practical one.
The moment you let AI write and create new notes, you have a new problem. How do you stop it from turning your vault into a landfill of machine -generated slop? Nick has a really elegant answer to this and it's one I've adopted after talking with him.
I would have a folder and it would be some version of, I would just call it AIOS like that. Then within that, Depending you might have a couple other folders, but I just have a place called history.
So anything that's created it I just know goes in a iOS history So it'd be like hey, I've just put the file in there and it would you know, it would give it the timestamp first So we could always just sort by date and then I would give it the title And so there's just this running, you know history of notes. It's created and This is crucial.
There's a very clear line of separation between my notes my thoughts and then where ai's notes and thoughts go and and that's the ai os folder only ai stuff goes in there and nowhere else that's the only place it can write to that i've given it instructions to write to with the one exception being if i say hey create a new idea note using you know the template and all that stuff that you know then it knows where to put the note okay but um yeah when when there's anything that's juicy and it's creating a new note because that we're researching something or in this case we're doing a dossier it knows just a ios history never have to think about it after that this history folder is the default save location of all artifacts created by claude so if you ever wonder where it is it's not in an anthropic server somewhere it is in your own owned and controlled but segmented part of your vault and then any of these other folders doesn't matter what they are but that's where all my other information all my other knowledge lives so there's a clear separation if i ever feel there's too much noise in the history folder especially like after years of these things that really have zero value
It's that easy to clean it. You just open it up and you say, I'm going to clean all these previous years. I'll put them in the deep storage, in the archive, and don't have to worry about them.
That is where the LLMs that you use, could be more than one, write and store all the texts that they generate, including any new notes that you specifically asked for. So you always know that 100 % of everything that AI has ever touched is contained and isolated in this kind of isolation chamber. And you know that everything outside of that isolation chamber is yours, is from you.
One of the most common questions that I get about AI and Obsidian is does the ai truly understand these double bracketed links and can it even create them for itself and are the ones that it creates valuable do they work now take this new project dossier note that you've created find every single term inside of it with a corresponding existing note and link it to that note and create a little map of content section if you're familiar with that term for these links.
Also, add a new section to the project dossier where you create a map of content, an MOC. Use all of the advice and recommendations issued by Nick Milo for that. Who dat?
That'd be insulting.
This is a great task. It likes this task. So verified.
Okay. Let me copy into the vault. 84 link targets.
So if we go back, I'm curious if the dossier has some links in it. Couldn't write directly to the vault folder. So I built the file in my workspace and copied it into via the shell.
Interesting. That's worryingly creative.
Okay, so they just appear. So these seem kind of like literal. There's not a lot of like creative interpretation.
It's like if I mention a book, it links to the book. If I mention quotes from a person, it links to those quotes. After about an hour, I got a little bit impatient.
Not with Nick, but with the output that we were seeing. We had asked AI to create a kind of project dossier summarizing what this project, the writing of my book, was about. And it was genuinely good.
It was accurate. It was well organized. But there wasn't really necessarily anything brand new.
There wasn't anything surprising that I didn't already know myself. So we decided to raise the bar and see if we could come up with a genuinely novel insight or observation. I'll think about it through two lenses maybe I'll stumble into something one would be through this project lens and then the other might just be more broadly so if I'm thinking through this project lens you already have a marketing plan I like that a lot and just depending on where that marketing plan goes and where it doesn't go One thing I'd be really curious about is based on everything from our existing audience, building a second brain and everything in between, the testimonials, the people who really loved it, what's going to resonate the most with them from this?
And that's probably part of there in some way. There are a few different things that AI can do uniquely well in a complex project like this, but one is going to be to analyze a huge amount of data. And this is like the imagination.
It's like, oh, well, Without this tool, we wouldn't be able to do that. And we were just trying to feel it out and just gut instinct it.
But this way, we can actually have a more data -driven sense of who are all these individuals in an aggregate? And are there avatars that we can put on them? Are there any surprises in the data?
Are there certain ways to reach out? And we will have more ideas based on what's already in your marketing plan.
i think the marketing side is is where ai can really come alive is since i have so much i have so much source material i just don't know how that source material can be sort of routed or translated to marketing efforts but that seems like a great task for ai because it's not wholesale invention it's really just mapping this goes to this we saw claude do the thing that i always see it do except this time we caught it on screen I gave Claude a short and incomplete list of emails that I wanted to write promoting my upcoming book.
And of course, as it always does, it treated that list as comprehensive. It treated the list as the entire truth, the entire reality about my book launch. Here's a question.
Is it over -indexing on one piece of information? It is. It is.
Which it does a lot. Which is great. You bring your own discernment.
This is not truth. This is not reality. And that's it.
That's, this is actually a great point because when this happens, we have to say, well, this is far from true, but is there anything useful? Yeah. This is overweighting on the fact that the list of email, email topics that I wanted to write on was, was very brief and incomplete.
Not. most or many of the emails will be about these things. They'll be about the annual review itself.
But those are kind of obvious, so I didn't list them. So this is such a funny thing. It's like, I think the tendency is AI always assumes that whatever information it has access to is all the information that exists.
You know what I mean? It's like... It's like every time you put in a prompt, you give it some working files, it's like a little universe, a closed universe that you've created.
And it's kind of understandable. It can only know what it knows. It only knows about things in that little pocket universe.
It doesn't know about all the other universes. So I think it's important to just think what you said. It's like, what is it over -indexing on?
And how can we build in rules in the future? Hey, AI is great, and a lot of people that are going to talk about it are going to say the amazing, amazing things it can do. But it does take tinkering and tweaking.
And so in this case, how can we prevent over -weighting, over -indexing of a blind spot? How can we, actually this was used in a different context recently, but how can we have generous assumptions? The bottom line and my takeaway from all this.
is that your vault isn't just your second brain anymore. It's also the AI's second brain. Humans have limited memory, but AI in most cases has no memory.
So it's even more in need of a second brain than you. So what a second brain means is changing. It's about creating this shared thinking environment between you, the AI, and potentially even other humans.
You both can create files. You both can read them. You can collaborate on them, but with some existing rules and guidelines and even boundaries between where is your private space and where is the AI space.
I think the real value of linking your AI is everyone is trying to learn about AI, but really we need to teach AI to learn about us. And when we can point it at our notes. and give it the translation layer that the manual like here's how to interpret and work with me it becomes really powerful and what's more it becomes future proof and it doesn't matter what the ai is because we own the core thinking back before ai moving to a new note -taking app or moving your platform used to be this gargantuan task that would take you know days or weeks take so much time and energy was so labor intensive and manual and error prone But that's really changed.
Today, moving your app or moving your platform is something that we all have to do all the time. You have to do it every time a new AI model comes out and it's better for your use case. Sometimes you have to do it because your LLM of choice is just down that day.
Sometimes you do it just because you wanna try out a new tool. You need your notes to be portable, to be easy and effortless and seamless, to just move from place to place to place. That I think is the single greatest benefit of Obsidian.
This idea of file over app, that you take your notes with you is a tremendously valuable principle that I hope we start to see with more and more different kinds of software.
The Hook
The bait, then the rug-pull.
Tiago Forte hands Nick Milo his real, in-use Obsidian vault, the one behind his new book, not a demo, and asks him to read it the way a doctor reads a scan. What Nick notices first sets up an hour-long walkthrough of how to back up, scope, and structure a note vault so any AI model can work inside it without doing damage.
Frameworks
Named ideas worth stealing.
04:05concept
Backup-before-access rule
If AI can read a file, assume it can eventually write to or delete it, even without explicit write permission, so always back up before pointing AI at real notes.
Steal forAny workflow where an AI tool is given folder or file access to real work product, not just note vaults.
08:00concept
Point-at-a-folder connection model
Of the three ways to connect AI to a vault (folder pointing, third-party plugins, or a CLI), pointing a Cowork/Projects session directly at a folder has the fewest moving parts and fewest ways to break.
Steal forDeciding how to grant any AI tool access to a personal knowledge base or codebase.
11:06concept
Folder as AI boundary
A folder is redefined from a human filing convenience into a literal permission boundary: the AI can't see or touch anything outside the folder or subfolder it's been pointed at.
Steal forStructuring any directory before granting an AI agent file access, to scope exactly what it can reach.
12:00list
Nick Milo's three-file translation layer
Me file (who you are)
Vault map file (how folders and structure work)
Skill map (what tasks/skills to run)
A small, portable set of three text files that tells any AI model, regardless of which one you're currently using, how to work with your notes and with you specifically.
Steal forAny Claude/ChatGPT project setup that needs to onboard a new model quickly without re-explaining context every session.
19:47concept
System-root file path trick
Instead of describing a note in general terms, copy its exact system-root file path and hand that to AI, guaranteeing it starts from the right file with no wasted tokens or ambiguity.
Steal forAny prompt where you need AI to reference one specific document with certainty.
20:49model
Sherpa system (style reverse-engineering)
Feed AI your 15-20 best existing documents in a category (maps of content, in this case) and have it reverse-engineer a style guide from them, then use a questionnaire skill to apply that guide going forward.
Steal forTeaching AI your voice, formatting taste, or structural preferences without trying to describe them abstractly.
26:44concept
AIOS / history folder quarantine
Every note AI creates gets timestamped and dropped into one dedicated folder, so 100% of AI-touched content stays isolated from a person's own notes and can be cleaned out or archived in bulk later.
Steal forKeeping AI-generated drafts, summaries, or scratch files separate from finished personal or client work.
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.
Nate Herk breaks down the four ways an AI operating system's context quietly goes wrong, then walks through the five habits that keep a growing second brain accurate instead of confidently wrong.
A full system for turning lazily-captured voice notes into a self-updating personal wiki — tagged, cross-linked, and rendered as interactive HTML graphs — using a handful of narrow agent skills.
A rant-format breakdown of why time and skill stopped being a business moat, and the four pivots (workflows, access, taste, enablement) that replace them.
00:28
00:59
02:00
02:27
02:49
03:25
04:22
04:50
05:16
05:39
06:16
06:40
07:16
07:42
08:11
08:33
09:21
09:50
10:33
12:00
12:12
13:03
13:08
14:16
14:35
15:12
15:27
16:09
17:09
17:32
18:01
18:17
18:59
19:28
20:21
20:49
21:13
22:14
22:43
23:16
24:06
24:25
24:58
25:26
25:58
26:49
27:17
27:50
28:08
28:49
29:34
29:58
30:39
31:18
31:44
32:07
32:30
33:00
33:17
33:42
34:11
35:22
36:17
36:46
37:12
25:03
19:12
24:27
14:35
11:53
Video of the Day21:07