Introduction

Hello everybody, my name is Marcin Angorowicz. Okay it's working, great.

So my name is Marcin, I'm a bit humbled to be in front of technical audience and talking about let's say technical stuff, not being a technical person because I'm an attorney.

Who I Am and What I Do

So, with nearly 15 years of experience, I specialize in new technologies law, intellectual property, data protection, cybersecurity, etc.

And before that, before my current company, which is mine, N77 Legal, I work, let's say, in leading Polish and international law firms. law firms previously it was one of the big four companies law firm and so and

basically about half year ago with my friend and co -founder we started our all two companies so one is law firm and seven seven legal and the second one is is a consulting company and 7 -7 works.

Why I Needed a “Second Brain” for Business

So, before this what I'm talking about basically is building second brain in the mold of Carpati's second brain in Obsidian and with Cowork etc. Now with Codex also.

Before that I was using AI tools for my work heavily. Before my current business i of course worked in bigger companies so most of the business stuff was covered for me

like marketing business development they basically like as people from sales they just pointed me to the guys and i went there and make new business and also of course working on the legal stuff

From Law Work to Running a Company

but when we started our company i found out quite soon that the legal stuff what i was doing for like 15 years it was quite easy the the hardest part harder part

was entire business stuff so basically now I didn't have like sales team marketing team and now it's all on me and my co -founder so a couple months a

couple months in the running our own business you know we are starting hitting the wall like backlog for the business staff pile up there was no no time for for so many so many things and we're like delivery like main products starting to ramp

up new clients etc the business stuff started to be really i we didn't have time for for like business development content creation uh working with crm etc so basically because of necessity i

Hitting the Wall: Business Operations Overload

Of course, I'm in the AI community, so I heard about this Obsidian stuff, this cowork stuff, and this stuff about second brains that can help you with running your life or business.

So of sheer necessity, I started to experiment. Let's say it was not experiment. I just knew that I needed to make something to help me with day -to -day tasks.

So, basically what I heard is you can create some empty files in Obsidian and you can connect it with Claude and this will create this great stuff that will be helping you.

So before that I was using AI mostly in the chat manner, I have like big prompt libraries etc. It was all great. great.

But managing entire context, files, chats, projects, GPTs, etc. started to be quite problematic on the time side.

Building the Obsidian + Claude System

So what I just basically did, I started building the second brain.

Starting Point: Prompts, Canonical Docs, and a Plan

So I started with basically perplexity prompt because I didn't trust like cloth to do it and chat GPT because I'm moving from chat GPT so I tried perplexity and so I described what I want and prompted it it gave me

entire plan how to how to do it so basically I started with I have some canonical documents like fact sheet about our companies our goals targets target, also second big file was ICP reports, so what are our ideal client profiles, etc.

And what I did after getting like a full instructions from Perplexity, I just dumped these files and also like chat GPT conversation, memory dump, etc., some templates for offers, etc. I just put it in cloud workspace. I set up the cloud and the cloud started analyzing that.

And from that, I started building and growing this second brain. And so I will show you how that works. So we were building this from the scratch basically. basically.

How the Vault Is Structured (Graph, YAML, Links)

Okay, so let's show the graph. This is how it looks right now.

This is of course file nodes and files in this entire scheme. They are connected with themselves. And there is animation, yeah.

And it basically started like that. I put some files, some canonic files, and started working with Cloud, with plan from Perplexity. We started adding

more and more this of MD files for for this entire system basically these are all MD files for example our strategy our brand voice our communication strategy etc etc I started building with Claude then Claude started connected with that we built in the files like YAML formatters starting to putting the wiki links together etc and categorizing entire stuff. So

basically how this this files works work is basically okay for example this is this is one of the document documents it got like tags, wiki links, dates categories etc and basically all these files what you just saw on the the map are built like that.

They are connecting with themselves. Especially, it is important for AI to navigate because all of this was created by AI.

After putting these canonical files, I didn't write by hand anything basically. Claude did all the stuff. So, it built that

kind of that files for example other other files there's content indexes etc and basically it it built entire entire systems entire system that looks like that right now so uh so this is uh this is this is it sorry i get back here okay yeah and and yeah sure Oh, yeah, sure.

These are different groups of, I don't want to, sorry for being a bit paranoid, because, you know, I work in this business for so long, so I got a bit paranoid. So, basically, this is like my catalog, there's my categories, for example, this is HQ, this is files that run entire configuration file for entire system.

This is like categories from files from the legal, for legal company, for consulting company, system files that there is full configuration here.

Tooling Choices, Privacy, and Local-First Constraints

Question. Can you consider a close notebook LLM or a very similar piece?

Yeah, but of course I've used, yeah? Did you use cohort for organizing all the files? Yeah, yeah, yeah.

Why Notebooks and Email Integrations Are Tricky in Legal

Like basically regarding other tools I always, I also using for example notebook LLM for some tasks but what i need but basically what it is this is tool that i built for purpose for purpose of helping me run business so for the legal task for example analyzing some regulation etc i have like dedicated tools i i've used also notebook lm this is right now it is not connected

for on the other system because of my specific of my job i cannot connect it cloth to my email because of client there's a ton of client privilege data there so basically I need to run different system in different instances on different even infrastructures so this is painful for me this is annoying but I cannot combine it together because of especially a ton of client privilege

so different tools from different stuff but what it this is this is basically only like first level of what I built and this is how mostly for doing business tasks tasks so this entire stuff is built on locally as MD files and this is of course Obsidian is just like interface

for that this entire categories I will I can show you I don't want to show of course too much because this is basically live you have entire entire category categories of of instructions you for example I have like configuration files my entire prompts system prompts chat prompts are not this

Technology-Agnostic by Design

is the source of truth for that. So basically this is because one of my main goals is that it was that this should be tool that is technology agnostic because I basically don't trust any given provider.

Secondly I want to have full control on that. I don't want to build it in like Notion etc.

I wanted to have it locally in my own pocket. Yeah. That's the case.

Why not have all this stuff locally? Sorry? Why not have it all fully locally?

What do you mean by that? Locally on your computer. It is locally on my computer. Yeah, it is locally on my computer.

So this is like a downside. But I know that you can build it, for example, build some VPS, do some mirror, for example, read -only,

and you can make it work from mobile because this is connected to your file system on your on your computer so in base level you cannot access it from for example from chat from the different computer if you don't have obsidian sync

you can install obsidian app you can purchase sync this is a paid paid service like syncing and you can put it in all other other computer but if you have for example Mac mini you can put it on Mac mini and access it from other

computers so but basically guys I this is of course it plans for expansion but this thing is working right now so basically I talking about how to build

Real-World Win: Generating Offers and Saving Time

it etc but what struck me the most when I started like month ago after three days after I spent entire days building it because I have little time and and the first first time that they really hit me it was like it was 10 minutes to midnight I wanted to stop my work but I I remember that I have one mail short

mail offer for my client to send so basically this was the problem that I hated doing the offers responses to RFPs if someone knows some if you work in companies when you have to put offers this is like annoying stuff you have to prepare describe what you want what you want to do etc of course you have

templates etc but what I did after just three days of building that we're putting this entire strategies, ICP list, etc. I just wrote like bring me offer with that type, with that tone, like the short one, short version, etc.

Generated it like in three minutes and for me it was like I would do it like half an hour for that and it made this offer in three minutes. So basically the problem that it was losing time for like 10 years, you can

like you are technical people mostly so you can laugh but in legal industry it's not very technological so you you do very much stuff by hand and so this is like problem that plaguing me for many years and after just like side effect of building this in three days it just resolved my problem and I have for many

The “3-Minute Offer” Moment

years so then i started of course expanding it and and at some point i like read the article about andrej karpaty and his llm wiki so i thought okay uh what is that maybe it will help my

my my my obsidian second brain so basically what i did i give to co -work link to the repo of karpaty uh co -work checked it and told me that basically i did i built this type of wiki just my is more more for the business side not for the research of coding because i basically don't code for that i

used i used only co -work and chat now i'm starting using a codex also for that so how many how much time do you have you're good you're good okay so so basically so basically about i i built like independent like 80 % of what Carparty thought of I think like this this this

Content Creation Demo and Playbooks

what I did my second brand is mostly adjusted for my needs so basically this is tool for business mostly from this business development work for content work for offers etc so now I will put it just small demo of course because this

is this is like I think there is like client not client the client are anonymized with codes etc but yeah yeah I'm too paranoid you know it's like 15 years in business so I'm a bit paranoid so I will just show it show you for like very basic thing which is even not finished like content generations so

Generating a LinkedIn Post Using Brand Voice Files

basically I can go to cloud just give me a second I can go to cloud so I can go to cloud and just give me a second I want to for example to for it yeah to prepare me LinkedIn post just basically this is in Polish so sorry but it means just create me yeah just create me new post so this this means just I need to create new post the LinkedIn post for AI act so just create me one just use your

style guides etc brand voice and this is starting to work because I of course used oh no it's sonnet it's very slow but of course hope there's internet oh it's starting loading tools it's not rocket science but for it yeah but I I don't want it to show it what it's showing me sorry where is it oh but if I don't want to wait I can use for example also codecs and do the same

because now I will show you what for me I'm sorry I'm getting to the end so but I will tell you what for in my for my opinion is the biggest benefit of of this entire system.

They are connected directly, they are operating on the on the files, local files, so yeah if I some of the computer codecs and cloud co -work is working on the files directly on the computer.

What is just happening because you see the cloud created of course the post and this is in Polish it is about like there is my and there is something like AI act and is coming into force and you have to prepare yourself if you are running company running AI but basically what it said this is

like final effect but what it did before just let me type it okay so I asking because you just saw the final effect, the post -generated, nothing fancy, nothing special, you can do it in basically every tool, but preparing that, you see, he is using some rules that are in the Obsidian base.

So basically, there is a file with brand voice, and there are, For example, forbidden marks, sentences, entire playbooks, anti -patterns, etc. Basically, he read these rules that are inside the base, like in policies, brand voice file, and used it to create this post.

How Agents Load Rules from the Vault

When you create your playbooks, do you actually just write out the instructions and the rules, or do you instruct Plot or Codex or Core to read other documents within, so embedded? Like if Core, if you miss something, check this entity file for these instructions.

Yeah, I didn't read anything. I just put in my goals, what it should do. For example, I'm writing you see that that's that's our rules that are used that files that system read to prepare this post

so basically there's this agents MD is only just converter from Claude Claude to To codex because the entire stuff is built on Like in code style. So there's like Claude MD files, etc

So to codex around that I just like prepared of course the codex prepared it just a short let's say the short file that just directs it to cloud infrastructure and translate it for him and it works codex works on directly on on base prepared and created by cloud so you see there's like brand voice there

is content style guide data governance because also there's some when it's producing some content of course it have to follow the data governance work etc so basically this is file I didn't wrote it there is no instruction in project directly this entire stuff it low it's it is loading it from the base and all these rules are from directly from the from the base so I don't need to prompt it and it just does stuff yep

Does it prompt its own iterations of the agent's file? Does it update the file itself? Yeah, yeah, of course. You can put it to prompt to change directly.

For example, when... Yeah, yeah, of course.

Operations: Memory, Logs, and Maintenance

So basically, for example, when I do scheduled tasks, because this entire base requires maintenance and controlling, it creates its own scheduled tasks.

The source of truth is in the vault, so no context is lost if I, for example, open a new account. I just need to put one general instruction to the

profile instructions, all the rest is getting directly from base. So going to the benefits of this entire system, because of course maybe for some of you this is like basic stuff, there's nothing special.

For me It is a very useful tool. It's helping me with my day -to -day work. It's making me much more efficient on the business side.

This is a basic version. There's no fully module for content creation, for analysis, etc. This is business development, basic content creations.

For example, when I choose if I should go to some event, I just ask it. It will check my entire strategy. If I, for example, don't like something, I just put in, oh, this element, I don't like it. We need to change, this rule, we need to change, adjust it.

And it just do it itself. So, basically, what is important after this first month of doing this, with that, I started to use Cloud. Because before that, I only used ChatGPT in, like, browser. So, with that, I started to learn using Cloud, Cloud Cowork.

I don't use code code code because I'm I don't know don't code but I start to use it codecs as a Example and it's really cool. So probably I start to go to back whatever

So basically, I don't write anything in the base. The system does is by by itself I Just I just check it for example for most more important stuff like strategic files, etc

I just check it read it. Sometimes it's okay. This is this is too bad but I observe the results.

I don't type myself. If something in outputs is... I don't like something in outputs, I just tell him to like change it, like put it

on the backlog that tomorrow we have to do something and tomorrow he reminds me and we, for example, change a row.

Managed Knowledge: Logging Every Session

The main important thing, this is from Carapati's side, because this is basically managed memory, managed knowledge and what it makes possible it is logs that this is this is logging in every session every day there's daily task who reads and all the logs from the

sessions take any important information for example if something is malfunctioning in the system etc every session everything is locked just if you started to do it just keep it like limit limit size of the of the operations log because when it's starting to go big, you just run through, burn through.

Yeah, I have only backup. This is backup in GitHub. Like, not in GitHub, in Git. Like, I don't know this stuff, so sorry.

But sometimes I just backup it manually. I do backup with Obsidian Sync and also with Git, et cetera. So, again, a bit paranoid.

So basically is when you when your operation log started to get big so you have to Trim it because it will burn to like two tokens like like absolute crazy But the this is this is most important logs like it's logging every operation is entire system It keeps learning because of that.

So this is like really cool. You just work on it day to day it remembers once a week, there's like entire scheduled tasks for for analysis of entire week and improving entire base and this is what

the once once a day like these daily days yeah and and next is the once a week like the big one the big conversation and etc I use scheduled tasking in in cohort yeah I started I started with perplexity I put like first I asked perplexity for plan and structure and first prompts and give me first prompts to generate for like a

Ymml formatting etc. I didn't know that stuff. So it built built me. I just thought it about us entire system

This is what important for me and all the rest is from the AI itself so so Sorry, yeah

Trimming Logs and Weekly Harmonization

When you say trim the logs, do you have a certain type of framework? Yeah, yeah, yeah. I learned it the hard way.

After four days, I would start one task in co -work, and it loaded only the files from the context, and it started compressing the session in one million token window. Is this a specific repo on GitHub?

No, I'm not backing the Git. It's only for backup. I'm doing entire logging in Obsidian itself, in MD5.

So there is limit for one entry, like this maximum 800 characters for one given time and if the file is bigger than some threshold, the maintenance schedule task will cut some of them from basically archives previous days. I want to make sure the quality stays the same.

Weekly schedule task. This schedule task is doing entire harmonization. It's checking Wiki links and checking everything is clear. I don't do it by hand.

There's too much of it. It's like 200 files right now. It is like crazy, crazy big.

Human-in-the-Loop: When I Review vs. Let It Run

So where is the, like one thing that I, I don't have the same system, nearly not as that, but question for you is quite often when I'll do a run, like what you're doing the sort of similar thing I'll have a output and I

put into the particular instructions that it will say instead of saying that will commit it will actually write for me do you want me to the following that I did is this good not like whatever do you have any sort of like human in the loop mechanism for this most of the stuff no like my human in this loop is

for example if I generate offer or post and I think that he is making like mistakes or using forbidden words etc I just check I ask him like why did we did you use this word there is a rule for that and for example he he outputs that okay but in this logical situation there is some ambiguity so I use that so

then we say I will just say because I'm working on for example pause I don't have the time to do like put it on the backlog to narrow this rule and next day we will do it.

But basically it helps that I'm like bit pedantic so like I will check this thing I will do it so for it to run flawlessly. Okay so it's

going to the end. Just remember I don't read entire files and And sometimes it makes mistakes, like small changes in the system. Like if you are a system architect, you probably know how it works.

Small changes, so if rules get a bit loose, it compounds in time. So basically, entire scheduled tasks, you have to maintain it with discipline or build many scheduled tasks for that.

Because if not, for example, after a couple of days, after a week, I found out that for example that the operations log was too big that with what was it was eating my context window Like only on loading face. So you need to

Yeah, yeah, yeah Yeah, this is this is Yeah, basically basically it does not rely on Project memory on system memory.

It's all structures. There are big big big files and rules that he is loading after start of every session. There are some rules and skills for it.

Different question. What I meant was, if you have, let's say, files that you save, do you have a system where,

okay, these will be good for a week. After a week, when I do the process once a week, I'm going to archive these into. Yeah.

It's doing by itself. AI just proposed me to that. It is versioning inside the files.

it's making it's having archive files with every every operation so basically yeah everything is marked out so or on JSON be bigger bigger for example my content base is in JSON basically but with with of course rendering and the

HTML dashboard but AI have to like work with that so most important things that you think is that the system has the best information for example is going to JSON, I don't need to see it. AI, Cloud or Codecs need to see it.

So basically, sorry to go into the end.

Limitations: Cost, Speed, and Context Windows

My biggest problem from my perspective is that sometimes it can be very costly. This is, you know, eating tokens is crazy. It's working slow.

The Codecs is answered. The Codecs is like super cool because it's very fast, especially if you put the fast mode, it can just do things in instant because cloud is like slow.

It's sometimes expensive especially on the building phase because MD files are eating tokens like crazy.

Use skills etc to optimize it but it can be expensive and it's sometimes slow but you know this is like just technical stuff.

Year from now with 1 million context window I just yeah just one thing for For me, if Sonnet right now with its speed had like one million context window in co -work,

and also was this cheap and fast, like have capabilities of Opus on the level of Sonnet, this is basically flawless. This is like perfect.

So just in the year, it will be great. Yeah?

Ideas for Cost Reduction and Local Alternatives

So to reduce the cost, let's say pretty similar to the obsidian, run it locally to let's say mark and index the pies. Yeah, but basically this was the simplest thing.

I just watched some YouTube videos. I think it's a good idea. I started with perplexity.

I know there's like GitHub stuff, et cetera. I just don't do it because I don't have time.

Are you vectorizing it? Sorry? Are you vectorizing it? What?

Vectorizing it. I don't know what that means. Anyway, I'll play it. Yeah.

So basically, main problems are technical, and the slow is sometimes expensive. But a year from now what is most important for I'm not non -technical I don't use yeah

listen these are not questions for me I don't know I just needed a tool this is this is like carpati rug I read that this is basically rug in some way I don't know I'm not technical this is for me this is like I needed a tool I made it tool it made itself basically with some discipline.

Conclusion: What Matters Most

Going back to Carpati, I think the most important thing is the entire concept, like knowledge base that is learning itself, managed by AI. This is the concept.

Guys, to conclude, in my opinion, this is proven. This is proven. It can be done right now.

Vendor Independence and Portability

Now it's a bit clunky because of like model is slow etc it's a yeah yeah basically basically the the beauty of the thing is it is for it was most important for me that it is technology agnostic because what you

can do for example i need to use like commercial version or everything because of business i do so basically i cannot buy the premium marks for consumer because it's only consumer terms so not data protection etc yeah so i need to use teams but there is no like this higher the premium seat it in teams is equal teams is like 5x not 20x like so this is problematic but with this you

can switch between accounts of course probably you cannot do it you shouldn't do it for example but your friend you can work on this base and this is working flawlessly if you hit the limit you can put you because you have we have also subscription in in codex in chat gpt in business so I can use it with with codecs with flawless flawless this is just works you can switch accounts you can do whatever entire memory is in the base yeah sorry

no I build like custom MCP in JS whatever perplexity yeah but but it just like provide output admit with code I just paste it in the file and it works so basically it's run like that MCP files for file system is the most

annoying thing of all because the beauty of the thing I learned is that chat and co -work are different things so basically it's not like there's one project that you can connect chat to this project chat cannot like work like co -work because co -work have direct access to your files on your on your

computer chat the problem with co -work is it's very heavy and it assumes you're gonna have a lot of work going in you're gonna start off with chat on a lot of these things which will lead to a project it will later lead to the project becoming something but co -work they were smart about it and you can

actually move those projects that you have now to on your computer yeah yeah I

just like that for example chat is better for research like deep research function you cannot have the same effect in co -work so basically you just do research in the chat then you just create handoff it

writes handoff to the inbox in the vault and you just switch to co -work and just paste the starting prompt and you can just keep working and also you can use this prompt in codecs for example this is beauty obviously this

is like technology agnostic I researched it like last week what am i vendor

are locked in here. So the biggest biggest problem is Obsidian. One element that locks me in to this to Obsidian is Wikilinks. This is only thing.

Without Wikilinks you can put it in every other software sorry yeah of course so it's basically this is technology agnostic, AI agnostic, software agnostic. You can put it on pen drive and have it with you in your pocket for me it is most important because i

Why This Is Already “Proven” in Practice

don't trust any big providers in like this is like critical knowledge and and one last things guys first this is proven in my opinion that this knowledge self self -learning knowledge basis can work and they are they're really really great and they are like 10x in my efficiency really and and the technical barriers are just technical it will fade in like next months on years so this is this is very cool now and for me this is the most important

that I'm not like tied to any given provider I can switch provider switch AI switch accounts and still works this is and also this is like pure knowledge you can of course read this files by yourself so entire procedures strategies

brand voices they are and written in the human form so you can work if for For example, internet fails, this files on your computer, you can still work. For me, this is super important.

So basically, sorry for, I think I stretched a bit. Sorry, thank you. Thank you very much, and let's get in touch.