I Turned Obsidian Into a Self-Learning Second Brain

0:00 Most people build a second brain, but I actually trained one, which means every time I use Claude, the next time it gets better. And the material it learns from is already on your computer, too, in a folder you never open. By the end of the video, you'll have two skills. One learns from the sessions you are currently working in, and the other reads every session you ever had and writes the fixes back to your vault. Yeah, so let's start but second brain actually is. Second brain is all of your context stored digitally. That's it. It can live on a cloud, it can be on your machine, and most people keep it on their folder in Obsidian vault, and they call it a second brain. And that's only half of it. The other half is this folder, which almost nobody opens it.

0:51 And that's where you have every conversation stored with Claude, and it's on your computer right now. Like, 99% of people don't know that it exists. When we talk about improving our second brain, what do we mean by that? Let's define it. One way to improve it is to add your context. Add more notes, more projects, more history, right? That's what everybody talks about. And the other one is playbooks. That's actually how you plus AI execute the work. And this video is about playbooks, and playbooks is the way you plus AI are working. Playbooks are simply put skills. And as you work with Claude, you have those conversation traces, and you can leverage this information to improve your second brain. And this video is specifically about the playbooks. Here is a memory layer, and the memory layer is just files on your computer. You open a Claude in a folder, and in that folder you have agent.md or claude.md. You have the skills, sort of playbooks, and here is the memory.md, which is just an index of all of your memories which are created automatically if you have auto memory turned on. Here the thing that you might notice that the memory layer, so this part gets ignored. And this is because Claude MD is read at the start and 500 lines later the model is no longer looking at it anymore. And if you put the same instruction into the skill and invoke that skill it works much better because it arrives in the most recent context. You know you have this feeling that you have done a big chunk of work with Claude and you tell it and you can tell that you teach it. You you taught it something along the way. And this is the moment when you want to capture this learning. But also it's the moment when the learning might just disappear if you close the session. Here is a skill and this one I use the most and that's going to help you to learn from your sessions. This skill code is a a respective skill. The respective skill runs exactly at this moment when you're coming to the end of your Claude session. It reads your current session, it pulls what actually changed, what are the corrections which you made, where Claude missteps, where we did the work, where is the work was redone, what actually worked and then it proposes edits to your actual setup. And you have a a table where you can approve or reject additions or you are in control.

3:21 So let's take an example right now. I was doing the outline for my video and here we got so many corrections and let's actually learn from the session. So you can see right now it's 400 tokens and it's a good moment to actually start a new session right now. What I do is I tell Claude, yeah, let's do a respective and learn from the current session. As simple as this. And here is the respective which Claude did just now. It was a bit proactive and it just went ahead and applied changes. That's fine.

3:54 You can control the behavior, but the main point is that those are real corrections which I have done, and Claude haven't them by default, right? You need to teach You need to tell Claude to use this retrospective skill to actually learn from the session. Next time I go through my video outlines, this never happens again. And it applies to every skill you run. Every session. This skill is a complete game-changer. One thing it might not be good at is that when you have a deep into the long context, well, here I had like 400K tokens, but if you are like go beyond that, maybe 6,000 tokens, the model actually forgets what happened at the start, so mistakes early on in the session do not always get reflected properly. Retrospective is something that you do only with a single section.

4:45 When you want to teach Claude something about the current work you are doing. It can't tell about what happened in this conversation. But, if you take a week work instead, this is where we have this new skill, which I call session intelligence. Well, the point is that well, we are doing right now most of our work in a loop with a large language model. And like in my scenario, it's 100%. All of the work is in those conversations, is in those sessions. So, and I don't see those patterns of my interaction with Claude or with any other provider.

5:19 and it could be that like I repeat the same thing over. And this pattern you could be missing. I'll build a second skill, which reads a week worth of session on my machine. It's very tunable. You can put it up to read conversations, and it produces one report, what you keep repeating, where the things keep breaking, how you actually work, and the findings become a skill or a rule. And in the end your setup gets actually better. Well, in a similar concept that you having this self-improving second brain. The way it works is that you have this folder dot Claude on your computer. It's a hidden folder and contains all of your conversations with Claude across all of the projects. Here is the projects folder and now here is my Artem wall.

6:04 This is where I work. I can go there and you can see those files and those files are actually my conversations with Claude all stored locally and you can see the time when it was edited. So and reflect the pattern of my conversations. So you can see they just go back in time and you can look at them and you can actually analyze them. And this is example of this file. It's has this weird formatting but it contains all of your messages. And the way I leverage this and I go about analyzing this is using sub agents because this whole conversations are there tremendous amount of conversations. I use Haiku sub agents to read the slices of the corpus.

6:47 We have here a bunch of sub agents and then after we have done with the analysis, we merge findings into the report. I run this one about once a week. Here is how this report looks like. I analyzed 91 sessions. It has overall my theme like where how I spend my time. It analyzes the work type. It gives me some amount of most useful skills. Actually you can see that retrospective is my most used skills here and here are the recurring patterns which happen over and over. This is just a summary but if you look into the road map here, you can see the actual analysis and ranking of all of the friction points which happen as you work. And here is example of a failure that happened. I use this tool called Annotator and typically that's how I annotate my HTML documents or markdown documents and here you can provide comments and once you send feedback what happened here is that feedback got lost and I basically wasted a lot of work. You can see that my actual message and here is the proposal how can we improve so this never happens again. And in the moment I didn't follow about that improvement and by running the skill it actually surfaced that. Another type which you can find here is skills. And here is a skills improvement backlog. You have all of your skills and it proposes improvements to the skills. Here is a planner data skill, here is my daily skill and this is entirely done by analyzing my conversations. And those are all the conversations which I had and this is analysis of this conversation. Now you can see what are the automation candidates for this conversation, what are the friction points here, what are the preferences we should capture and also the quotes from this conversation. It's really going into each of those, diving very deep and learning from them. What does this buy you? If you leverage all of your sessions or all of your conversations, the next piece of work which you're going to be doing is going to be much better. It's going to be less friction for you to go through your work. You actually stop re-explaining yourself, you stop repeating yourself. You're actually training your second brain to work the way you work. It's not the other way around. And actually your work becomes more enjoyable and this is the part you don't expect. If you keep repeating yourself a cloud, this is for you. And I want you to grab retrospective skill in the description and run it on one session. And that's the whole first step. And this is what I run every day. And if you want retrospective skill and the whole system built, this is a 30-days a journey second brain. Link is below and see you in the next one.