Obsidian + Jev é absurdo! Segundo cérebro em tempo real (veja como)

Author: Matheus Battisti - Hora de Codar · Watch on YouTube

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: Obsidian + Jev é absurdo! Segundo cérebro em tempo real (veja como) · Channel: Matheus Battisti - Hora de Codar · 24:35 · Watch on YouTube

Topics: AI agents, Task and workflow automation, AI developer tools, RAG, search and information retrieval

Summary

The video demonstrates how to combine Jev with Obsidian to organize notes, create links between them, and find relevant information more efficiently. The creator builds an Obsidian plugin that uses Jev to suggest folders, note types, and tags, while leaving uncertain decisions for human review. A second feature, the “librarian,” filters vault notes before Claude Code answers a question, with a comparison showing a shorter, more focused response and lower token use in the demonstrated case. The video emphasizes that Jev supports existing AI workflows rather than replacing them.

Key points

Jev is presented as a fast, low-cost decision-making tool that can support Obsidian workflows through an AI agent such as Claude Code.
The “bookshelf” plugin reviews notes in an inbox and suggests destinations, note types, and tags based on options already used in the vault.
The plugin offers a preview plan before applying changes, an undo function, and a human-review path for uncertain suggestions.
Jev evaluates note titles and beginnings, returning confidence scores that the plugin uses to decide whether to apply suggestions or request review.
After organization, the plugin can evaluate note pairs and add links for related or complementary notes.
The “librarian” sends a question and note details to Jev so it can identify a limited set of relevant notes for Claude Code.
The librarian can return up to eight notes, prioritizing direct answers and then contextual help, and can report when no notes qualify.
In the demonstrated comparison, using the librarian produced a more focused answer, used fewer tokens, and was reported as nine seconds faster.
The creator cautions that automatic organization should only be enabled once the workflow has been verified, since automation could otherwise cause problems.

Chapters

Start	Chapter	Summary
00:00	Overview: Jev and Obsidian	The creator introduces a workflow for using Jev to organize an Obsidian vault and find information more effectively.
02:05	Organization rules and human review	Jev evaluates each note using its title and opening text, while uncertain cases are reserved for a person to decide.
03:36	Building the Obsidian plugin	The creator previews the workflow and begins building a plugin called “bookshelf” to organize notes in an inbox.
08:15	Connecting Jev to note organization	The plugin’s initial interface is tested, then Jev is connected to suggest folders, types, and tags with confidence thresholds.
12:24	Testing and refining the organization workflow	The creator fixes when analysis runs, configures the Jev key, tests the suggestions, and adjusts the displayed timing.
15:40	Creating links between notes	A second plugin feature evaluates note pairs and adds links between related notes, with undo support.
17:44	Building the librarian	The creator sets up a librarian that uses Jev to identify which vault notes are relevant to a question before Claude Code reads them.
20:20	Comparing librarian and direct search	The creator compares Claude Code answering with and without the librarian, reporting a more focused result and fewer tokens with the librarian.
23:01	Results and closing remarks	The creator discusses potential automation, warns that it should be verified first, and summarizes Jev’s role in saving time and supporting better decisions.

Quotes

If it’s not sure, the note is set aside for you to decide. — 02:05

I can see what it’s going to do beforehand — 07:14

If none pass, it clearly gives up instead of developing anything. — 19:18

It was faster, had token savings, and the answer was more accurate, more straight to the point — 23:01

It’s a tool you’ll be able to use, right? To save money and gain time — 24:04

Concepts and entities

Obsidian, Jev, Claude Code, TypeSafe, bookshelf, vault, bibliotecário

Full transcript (original language)

00:00 Guys, Jev with Obsidian is amazing; it really improves the experience. In this video, I want to show some ways I’m using these technologies together to have a more organized, faster, and truly useful second brain. I’ll show you practical strategies so you can implement them too and merge these technologies. So, come on in, let’s go. So, guys, check out what this video will consist of, okay? I’ll show you in a few steps how I integrate Jev into

00:31 Obsidian. I’m using Cloud Code here as an example, but it works with any AI that can connect to and understand Obsidian. The initial idea is that Jev has a very fast response and can make decisions. So, the organization part of the notes we have in Obsidian, in the vaults, will be much faster and more accurate. We’ll use it right away, and it will come in and breeze through everything, organizing everything we have placed in our vault. So here, guys

01:02 , I’m going to break it into a few parts, as I told you. And the first one I really like is the organization part. I’ll call it “Distante” here just so we can give the project a name, right? But that’s up to you. It will basically be a plugin inside Obsidian, right? That will read the notes there and pass them through Jev to organize everything, tag them, and link them— making connections between them. So, it will be much better. I created a space there in Obsidian just to show you how this is all done and how it looks in

01:32 the end, okay? And there’s another one I’ll call “Librarian,” which will be a skill linked to the vault along with Cloud Code or whatever you use, which will use Jev to find exactly what we are searching for. So this helps save time and tokens because Jev will decide exactly what we want, since it’s a decision-making AI, okay? So these are the two main points I want to touch on, but I’m sure it will open your minds and show other possibilities , especially if you use Obsidian a lot, right? So, guys, how did I organize

02:05 this here? Jev will decide for each note, right? What will it decide? Well, it will read the title and the beginning of the note and choose between the folders and labels we already use. If it’s not sure, the note is set aside for you to decide. So, it will perform this cleanup, or this organization, based on these references here, and it will still leave what remains, let’s say, ambiguous, which it can’t quite understand, for a human, which is the best situation, because, I believe that better than an AI making

02:35 the wrong decisions is us being able to intervene and choose the right way. So, we will have some examples here, just so you can understand how it will make these decisions later. So, if it can’t understand how it can connect, which label it will give, what the references are, it will send it to a human to review, which is us, right? And then we will be able to make the best decision. What do we gain from this, guys? With Jev, we can have much more precision,

03:06 much more agility, and not depend on an AI, an LLM, which for this case would be very slow. So, since Jeff is fast, it’s built for this, to make decisions and be very cheap, it will be done with great mastery and will help us do this part in a more automated way, right? You get a job very well done in seconds , right? Sometimes it won’t even take a second, depending on how our VTS is organized there. All right? So, I hope this provided a general introduction and helped you understand the

03:36 connection, the connection point of JEV , but let’s see it in practice here because it will be much more interesting since we will see things happening. Let’s go. Before we actually go to the practice, guys, I want to tell you that we are producing our Jeff training, which will launch next week. It is a straight-to-the-point course where I will show you the main features of JEV, how to integrate it into a project, and how to create projects that really have value with JEV, right? It’s no use just applying it like crazy , putting things into little games that you will never use in real life, right? I will show you how to do this

04:07 professionally to extract value from the technology and put it into production, right? This deployment process is also very interesting so we can also separate the environments, right? Development and production environments. So, all of this will be in this training, and the pre-registration link for the founding group, as I always do here, is in the pinned comment. So, pre-register now, and when it launches, I’ll let you know so you can get this special discount price while using AI professionally. So , let’s go again, right? Let’s go. Look , I created a very fictional vault here

04:38 , right? Everything is disconnected in it, right? But it has everything a vault would have, right? Here I have some notes on AI tools, here I have some study notes, here I have some journal notes, right? Something we would do on a daily basis using Obsidian, okay? So, what happens? Everything is disorganized, they aren’t connected, and AI will act precisely on this part at the beginning. So I have

05:11 my VS Code open here with the second brain folder. So I’ll keep everything close, the project and my Obsidian vault, okay? And here I will start the “bookshelf,” which is what will organize things—that’s how you will do it too, okay? If you want to follow my AI with Obsidian idea, so I’ll start with Claude, just to remind you guys, I took some time to reach this result here. I’m giving the prompts straight

05:42 to the point, more generalized so you can take advantage of them, but you might need to make some changes to suit your scenario, okay? So here I’m going to create the Obsidian plugin, it’s going to be the “bookshelf.” I’ll send it like this to Claude for any AI. Create an Obsidian plugin here in this “bookshelf” folder, the vault is in the “second brain” folder, the vault is next to this one, it’s here in the structure, and the plugin code doesn’t go inside it. Whenever you generate a new version of the plugin, copy the files into the vault’s Obsidian plugins

06:13 folder just so I can reload and test it , to keep it organized so we can maintain version control. The “ bookshelf” organizes the vault’s inbox folder, okay? So we have an inbox folder here. For each note, it will suggest the destination folder, the note type, and the tags, always choosing from the folders and tags that already exist in the vault. This first stage is just the interface with dummy suggestions, without AI, that is, fake data just to create this initial screen so you can visually see what I’m

06:43 talking about, okay? The interface panel on the right side of Obsidian, opened by an icon in the top left bar, shows how many notes are in the inbox, right? That folder over there that’s set as the inbox. Below, a list of the notes, each with the title, the suggested folder, the type, tags, and the confidence indicator, which is Jeff , right? Green when the suggestion is secure, yellow when it goes to a folder to review. Clicking on a note opens it in the editor. There are two buttons: view plan, which just shows what would be done, and organize, which applies it . So I don’t have to rely blindly on it

07:14 ; I can see what it’s going to do beforehand, which is also pretty cool, right? It makes us feel more secure than just executing it and having everything get all messed up afterward. Let’s also add an undo button, which is interesting, as it reverts the entire last organization. So, if I organize, I can “disorganize,” let’s say, okay? Let’s set up the buttons; it still won’t move any notes yet, right? Everything is still fictional. And a plugin settings tab with a field for the TypeSafe key where we’ll put the JV key later, okay? And an indicator showing if the key is filled in. I’m drawing the icon, the symbols, and SVG by hand. It’s just to make it look a

07:45 bit nicer here, okay? No emojis, no icon library, just to look more professional. So, it’s creating it here , guys, and I’ll show you how it looks in a bit. We’ll just need the TypeSafe key. TypeSafe is the company that created JV, right? So you’ll log in here and at the API keys section, you’ll create your key, okay? I’ll create mine real quick. Then you copy the key somewhere safe, as we’re going to use it soon, guys. Let’s just wait for it to deliver it to us here, okay? So look at this, guys, it’s finished here, okay? And what happens next? The plugin will go directly into Obsidian

08:15 in this case, right? And it said that to test it, we need to reload Obsidian, okay? And we have to disable the plugin restriction there, right? When we open it, it asks us, and we just click there , and everything will be fine, okay? It mentioned what’s there and such, but it’s easier to just see it, right? Oh, it already appeared here, okay? The shelf is there, and we already have everything we need, right? Visually here we can click on view plan, and it shows, you know, all dummy data for now

08:47 . You click organize to perform the organization, and undo to return to the normal state. So it took all the notes here and everything else, tagged them, and it’s already working. This is the initial part, you know, having this linked in Obsidian for us. Now we need to integrate Jev to simplify this note organization, in other words, get Jev running here in Obsidian, via the cloud intermediary, you know, and make the magic happen. So, I’ll come over here

09:17 , and we’ll send a prompt here, okay? Now Jev enters the skills of Jev that is already installed in Cloud Code. Hey guys, there’s a Jev playlist here on the channel, okay? I won’t explain every detail in every video, or it gets a bit tedious, but check out the playlist, it’s easy to find here on the channel. The recent ones have a lot of Jev videos, and I explain a bit more about the skill there, you know? But basically, it will come here to home, okay? You will copy this prompt here and send it to your agents. Then it will install the skill, okay? Even if

09:47 it’s set to cloud here, right? But you can come here, look, edit setup, and then you can add other agents and you can install any code agent, okay? This skill is fundamental for the code agent , for the AI we’re going to use, to know how to use Jev, okay? So, coming back here, for each note in the input, send the title and the beginning of the note to Jev with these questions in a single call, right? So, it will also use parallelization, which is a Jev feature, okay? Which folder does it go to? Among the existing folders in the vault, with the “none apply” option. What is the note type I described here,

10:19 that’s why I’m saying, it has to match your vault too, you know? Idea, task, reference, study, meeting, or diary for each of the most used tags in the vault, if it matches the note. Then I set it here so the decision remains in the code, apply the tags that pass 70%, so Jev returns a reliability score , you know, from 0 to 100. So anything above 70 is ultra-reliable, it can go ahead. If the folder confidence is below 60%, then we’ll have a human review. Organize, move the note to the folder, and write the type and tags at

10:50 the top, without changing the text, right? So it will just edit it so that it is actually organized. Save everything that was done for the undo to work at the top of the panel. Show how many notes were organized, how many were for review, how long Jeev took in total, and how much it cost in Reais. Because I’m doing this demonstration for you here, right? So I increased the profit, but that’s not needed. Using the input tokens of what Jeev returns. The price I put here, right, is 0.04 cents per million tokens, plus a dollar exchange rate that is editable in the

11:22 settings, okay? So, just to be very precise. So it’s creating the configuration here and we’ll see it in a bit, guys. That’s why we need the Jeev skill, right? Because it runs tests, right, to check and see if it’s working. It does, look, testing Jeev decisions. So all of this will help us have better quality in organizing our notes and the general work in Obsidian. So it’s essential to have that skill installed, okay? So guys, look, it

11:53 brought me the result here, okay? And it said it analyzed the notes and so on , right? And so how does it work now, right? One call per note, decision in the code, organize, move the note, write type/tags, undo, return. And we’ll have the functional buttons now, okay? Here, guys, I just found an issue where it says the analysis starts by itself. Actually, no, I have to first click on the “view plan” button to understand what happens, okay? So, I’m

12:24 going to add this correction here for it. So, I’ll put it like this. Jeev can be called when I click on “view plan,” never by itself when I open the panel, okay? I’ll ask it to fix this because otherwise it would do things without us asking, which is bad, right? I thought I had specified that, but I don’t know if it misunderstood at some point. So I’ll fix this just to be in line with what I was expecting, okay? Look, now it’s fixed. So let’s take a look at how it all turned out, okay? If we go to

12:57 Obsidian, you can press Ctrl + P, reload without saving, and it will open up here, updated. So there’s no JV key yet. I’ll open the settings. And here is the Type safe key that I mentioned to you, right? So I’ll put mine in, look. I’ve filled in the key. Great. This is the dollar exchange rate, right ? Just so we have the exact value, you know? It wouldn’t be necessary. And now I’ll reload again, just to be sure, and click on “view plan” to understand what it will change, right? So now we have

13:30 the real data here, right? Without those fake analyses that were there before and such. So let’s go. Look, this was the JV analysis, right? Eh, I think it miscounted, guys. It wasn’t 20 seconds, I think it was only 2 seconds. But that’s okay. We had this cost here, I also think it’s wrong, okay? I think everything was much faster and cheaper, right? Maybe it counted by Claude’s tokens. I’ll just try to fix this because it seems way off base. You

14:04 could follow along there, it was very fast, right? And it organized everything here, okay? It already managed to determine the folder, it already managed to categorize, right? And some still have doubts, right? But it was expected that this would happen. But if we were to organize it here, look, see that it has already changed, right? And so we have a few here to review and such, but I can undo it and go back to how it was, just so we can also understand that everything works.

14:36 Well, before continuing, guys, I’ll just put in an adjustment prompt here, because the values and time really are wrong, okay? And I hope to fix it this way. I won’t insist too much on this if it doesn’t work again. As you saw there , right? We saw that it was very fast, and that gas that appeared there, I think it was the gas I’ve had on my entire account using JV since it launched, okay? But let’s see if it adjusts, otherwise we’ll go to the next step, which would be connecting them, which is also in this plugin of ours, okay? Look, guys. So, it returned.

15:08 Let’s take a look, see if it corrects it. Quickly here, view plan and let’s see it organize. Look, now that’s more like it, right? That was the real time, and the cost might still be inflated, guys. But that’s okay, the time is finally realistic now. Almost 3 seconds to organize all these notes, right? I’ll move on then to the other stage, which is the linking stage. So, the linking stage is a bit simpler. Now, the links. Add a link notes button to

15:40 the panel. For each note in the vault, the code selects up to 15 candidates, notes from the same folder or that share a tag. Then ask Jeff for each pair if both notes cover the same subject or if one complements the other . Send only the title and the beginning of each, right? So he can understand the context. Pairs above 75%become links. At the end of the note, a “ related” section with links to the other notes, without duplicates, if they exist. The undo function also undoes the links. I put it here, right? It shows in the panel how many links

16:12 were created and how many pairs Jeff evaluated. So this would be the second stage. Over there in the organization part. First we would organize everything into folders, and then make the links, since they would already be organized in their ideal place. Of course, right? There would also be human review to ensure everything is in its right place without him guessing incorrectly. So, guys, he delivered it to me here, okay? He said he ran the tests and everything else, right? They

16:43 found 387 pairs, right? So he looked at the 60 notes, and he said the button is there. So let’s check here. I’ll reload it too, look. I have to link the notes. So I’ll click here to see the plan. I organize them, and I can link them now, right, the notes with the other notes already organized, right? Look, Jeff evaluated it there and made the links.

17:14 So now I also have the notes connected, right, related practically automatically, right, guys. So it’s much better to get things done this way , right? Even if it’s not a brand-new vault like this one, if it’s yours that you’ve already added notes to, right? It’s a work in progress, you can still get new links, have better organization , it can still be improved. And now, guys, I’m going to create the librarian

17:44 , because I can use Claude Code to query what’s inside our Obsidian, right? I could do that in a very simple way. It can go there, evaluate, and bring back the answer, right? Many people use this, right? Claude Code as an AI agent for their second brain. But I want to implement JV there to help me find information even more easily. So, this way I can use the power of decision to make life easier for Claude Code and save me time and tokens. So,

18:16 let’s go. I’ll put it to him like this, look. Now, the second role for Jeff: librarian for the Claude Code I use inside the vault. Today, when I ask something about my notes, it goes around opening file by file, and it’s like us creating projects, right, systems with Claude, it reads everything and so on, gets context, and only then does it change or add things. I don’t want it like that; I want it to ask the librarian first and read only what matters. So the librarian will be inside the second brain part as a command that takes a question in

18:46 Portuguese and returns only the right notes. It uses the same TypeSafe key, right? So it will use JV for this, and it decides in this way here. For each note in the vault, send the question to JV along with the title, folder, tags, and the beginning of the note. With two questions. Does this note help answer the question? What kind of help does it provide? Answers directly, provides context, just mentions the subject, or doesn’t help? So JV here will give me more information about what I want to know by hitting the AI and answering me exactly these things here. The calls go in parallel, about eight at a time. So we can do this, right? There is

19:18 parallelization of F. It returns at most eight notes that pass 50%, first those that answer directly and then those that provide context, each with the folder, percentage, and type of help. If none pass, it clearly gives up instead of developing anything. So we also have JV being honest, right? Not inventing, hallucinating. In the terminal, the result appears with a short table below, how many notes it evaluated, how many it discarded, and the actual time and actual cost. Just so we know here too, but it is disposable in the real system. Each query also appears in Obsidian. Add a

19:50 librarian tab to the bookshelf panel with the latest queries, the question asked, and the notes chosen, with a click to open each one. So I can have a history of what was asked. Finally, create a skill in the vault, okay? So it will be based on this skill here for Claude, right? Write in the vault’s cloud.md how Claude should work. For any question about my notes, it runs the librarian first. So Claude will enter this librarian structure to be able to answer things, okay? So it will have this JV filter beforehand which

20:20 will make Claude much more accurate. I’ll show you in practice here, look. I’ll let it finish here. Well guys, then it delivers the librarian here. If we go to Obsidian, it will be here, look. Here it will map the questions we ask it, okay? So I’ll do a test with and without, just so we can compare if it turned out cool, okay? I’ll open two terminals for each session, right? And the first one, for the first one I’ll ask it not to use the librarian and

20:52 send a question. So it will be the following: don’t use the librarian or its skill in this conversation. Search in your own Claude way. Which video ideas about cloud code did I write down ? And I haven’t recorded yet, give me the list with the name of each note. So , all right, right? Here, how would I use the second brain with the structure we made for the librarian? No need to say anything, it will use the skill, right? The librarian already noted it here, okay? Here it didn’t use it. And we can compare the time aspect, okay? Also see if the context was less

21:23 cluttered. So here, why isn’t the answer immediate? Because Jeev is acting, right? But Claude still acts to be able to formulate the answer. Jever only does the decision part, right? As I was commenting, let’s suppose, right? Even if it takes longer, right? We would still have the accuracy of the answer, right? And here guys, look, it answered first, right? It was a bit first, it wasn’t that much of a difference, right? It was 44, it was 9 seconds less, okay? And even if it were

21:56 a little more, right? We would most likely have a more accurate answer because the JV is inserted there deciding what each thing really is, right? So let’s see here if we were to look for those already recorded or published. There’s Work, cloud code from scratch, cloud MD. Let me see here . You can see, right, that the question was: “Give me the list of notes that I haven’t recorded yet?” So, that was exactly what the one with Jeev did.

22:28 Because it answered me only what I asked for here. He brought everything, the recordings and such. So there’s that too, right? But here the answer is messier, isn’t it? The one with just cloud code. Well, if we look at context here, it used cloud code, it’s right here with 64,000 tokens out of 1 million context tokens. Let me see here . This one also has fewer tokens, see? We also had some token savings. 5,000 tokens just in one go, right? Probably

23:01 with usage, the savings would be even greater, right? So you can see it, right? It was faster, had token savings , and the answer was more accurate, more straight to the point, okay? Guys, that was the demonstration I wanted to show you, okay? What can we do later? Improve it even more, automate this organization, like that first wrong version, in quotes, that I went and fixed, let it run freely, but then we would need to be sure everything is working as we want, right? Otherwise,

23:32 this automation could backfire; it would let JV always receive the invoice , classify it, trigger it, organize it, and that’s it, right? You wouldn’t need to worry about anything, just drop the invoices into folder X and it would do the rest automatically and immediately, virtually for free, right? Without involving another AI in the process. Got it? Well, guys, a reminder for you, right? Don’t forget to pre-register for our Jeff training; if you liked what you saw here, there will be much more inside. So, I’m waiting for you there,

24:04 okay? The link will be here in the pinned comment. And now I want to know your opinion, right? If you like the use of it, do you think it will be interesting for your workflow and if you understand the purpose of JEV, right? As I’ve been saying here, a lot of people in the comments have been saying: “Oh, this Jev is nonsense.” But they didn’t really understand this part about the savings, which is an extra benefit, right? It fits into the flow; it’s not meant to replace anything. It’s a tool you’ll be able to use, right? To save money and gain time, that’s the main thing, right? To make

24:35 better decisions, because that’s what it was made for, okay? So that’s it, guys.