329 | The $0 Knowledge Graph: My AI Knows My Entire Business, So I Do Less and Get More

0:00Hello, and welcome to the Leveraging AI podcast, the podcast that shares practical ethical ways to leverage AI to improve efficiency, grow your business, and advance your career. This is Isar Maitis, your host, and I've got an awesome episode for you today. And this episode is going to take you from, yes, I'm using AI, I'm doing cool things, to having superpowers on how AI knows everything about your business, about your clients, about your projects, about the things you're working on, and will be able to really become a superpower for everything that you want to do. and we'll be able to give you much better answers, much better solutions, much better capabilities in significantly less work.

0:40So if that sounds something exciting to you, stick around. This is our topic for today. The topic is how to build a knowledge graph on top of all your knowledge and make it accessible to your AI tool so it can give you answers with the right context without having to give it the context every single time. So before we talk about what exactly is a knowledge graph, how to build one, and I'm going to take you through all the steps, how mine looks like, what components does it have in my universe, what you can build, and so on, I want to talk about the why, which is the most important component. The most important part about AI in order for it to give you better answers is context.

1:19AI that has the right context will give you much, much better results. Now, the reality is in order to build a solution in AI that will go from, oh, I need to tell it everything about myself every single time to I need to tell it very, very little and it can act extremely well in an agentic way, develop things and solutions for me without me having to invest almost any effort. There are a few different steps that everybody went through. The first thing was prompt engineering. So we all learned to tell AI what role does it take and then what is our background and to attach the right files and to explain everything that we're doing.

1:55And it was really necessary because without it, AI did not know any of that information. The second step after that was context engineering. Everybody started talking about how do you build an environment in which AI has full access to your context? The next level was harness engineering. How do you build an environment around the environment that comes out of the box? that OpenAI or Anthropic or SpaceX gives you. So how do you build more of the system, the ecosystem around it that's connected to everything that you need in order to provide it? And then there's different levels after that of agentic levels engineering or process engineering that actually tells AI how you work, how to work with you in your environment, connected to all the things that we said before.

2:39As you grow through these levels and build a better system, you will need less and less prompt engineering as an example. And I will show you later on that I can use really shitty, short, one-line prompts that gives me incredibly detailed, accurate results. And the results could be just an answer, or it could be a whole automation that runs components of my business because all the other things are in place and it has a lot of information and I do not need to explain it to the AI every single time. So this progress from prompt engineering all the way to process engineering and everything in between is something that anybody who's chasing better and better AI capabilities has went through.

3:19And I'm going to share with you a very clear step-by-step guide on how you can build your solution to be able to work in that level of effectiveness. Before we dive into why you want to do this, let me show you that this episode is brought to you by the Multi-Agent Orchestration Course. This is a course that we have been teaching extremely successfully since April. We already have probably a few hundreds of people who took the course. The course is taking you from whatever your knowledge is right now to being able to build advanced multiple agents connected to one another, developing extremely complex and sophisticated business processes with agents.

3:57The next cohort that is available that you can sign up for starts in November. And if you want to learn more and sign up, you should do this right now because these courses are filling up very, very fast. And if you want to know how to do this, there's a link in the show notes that you can click and continue and learn more about it and join us, hopefully, if you're interested in starting next year or finishing this year with a lot more of your business and of your life automated by AI. Now back to why you want a knowledge graph and why do you want to build a layer that understands all your context that AI has access to.

4:33First of all, as I mentioned, context is the most important thing. If you're just running from one chat to the other, well, your context dies every time that you are finishing a conversation and starting a new one. Now, I know what some of you are thinking. I'm like, oh, that's not true. Claude and Chachapiti now have a built-in memory. And you're correct. They have a built-in memory. They know a little bit about you and a little bit about your business, but they don't know it down to the granular level of every single step, of every single nuance, of every single project or client or anything that you're working on.

5:02And that level is required in order to get you to the outcomes that you're trying to get that are specific actual solutions you can use in your business. So continuous context on everything that you're doing is the very first reason. Now, I know what other people are thinking like, oh, I have a solution for this. I ask AI to write me a briefing at the end of each chat that I can give to the next chat, like a handshake document, so I can continue my conversations. And this is how I started as well. But then I understood that is very painful and token heavy. And I have to do this at the end of every chat.

5:35And then I continued from there. The other thing is you keep on making decisions that are based on things that you're learning and so on. And you don't want to remake these decisions. If you decided you want to use a specific tool for a specific purpose, you want AI to know that. You don't want to figure it out together every single time. Even if you've made that decision four months ago or six months ago, If it's still relevant, you don't want to go through the entire decision-making process as well. Another big reason is mistakes. We all make mistakes. The AI makes mistakes. The processes are not perfect.

6:07You do not want to repeat the mistakes. You want to be able to avoid every mistake that you've done before. And this is another thing that you get from doing all of that. The last component that is maybe the most important is pure efficiency and capabilities that are at a completely different scale. As I mentioned, I can voice type and I voice type everything right now. I don't type anything anymore. I can voice type one sentence and get a fully working solution for something or a great beginning to a brainstorming session that otherwise will take me a few minutes and a few pages of data that I need to aggregate and bring together for different sources.

6:49in order for the AI to provide me the capabilities that I need. So this is the why, right? I can get significantly better results of any kind that you can imagine with significantly less effort, which is the whole promise of AI, right? Is you want to get better outputs in less investment. And this allows you to go to the next level. So let's talk about the five different components that I have in my system that enables this magical capability. So the concept of a knowledge graph sounds like something an engineer is required and you need databases and stuff like that. And to be fair, that was my original plan.

7:26And I'm going to talk more about this in the end of what was my original plan and why I did not do it in the end. But right now, all of my solution that works extremely well at my scale, which is hundreds of files and potentially lower thousands of files, is all based on text files that the AI just knows when and how to read. And as I mentioned, it has five components. So the first one is I have one main folder. That one main folder, all my AI is connected to it. And those of you who listened to the previous episode know that I can connect any AI that I want. Most of the time I run in Claude, but I can run in Chachypt or Grok or anything else that knows how to read and write files on my computer.

8:07So I have one centralized folder and then I have subfolders under each and every one of those based on different areas of my business and of my life. As an example, I have a personal one that doesn't get backed up and so on. I have clients one that is not being backed up and is addressed under an NDA with each and every one of them in its own separate universe. but then have three that the AI always has access to based on different aspects of my business. You can arrange this however you want. Now, under those five categories, I have multiple what I call projects. And I know the name might be confusing because there are projects in ChatGPT and project in Cloud and so on.

8:47I just stopped using projects in ChatGPT and in Cloud because the setup that I'm teaching right now is just significantly more efficient. So for every new thing that I want to work on, I tell AI that I want to work on a new thing, and it creates a new folder on my computer under one of those five categories that I mentioned before. Right now, I have 53 of these projects across my entire workspace. Now, for each and every one of these projects, the AI generates two files. One of it is called, or I call it, the evergreen document. It's basically a briefing document of what is that project, what decisions we made, how many conversations we had, what were the high-level topics of each and every one of these conversations, what lessons we've learned as we were working on this particular project, which prevented from repeating these mistakes, and so on.

9:35The other file is a task registry. It is a CSV file that basically holds every single thing we need to do, and the AI keeps track of both of those. I don't read them, and I don't write them, but the AI can read and write these files and stay completely up to date on exactly what is the current status? What needs to be done next? What are blockers for other things that we have to do first before we continue? And all these kinds of things are all in these files. Now, diving down to the details of these files is going to take way too long, but you can build it in whatever way is going to work for you.

10:07The additional file that exists at the top level, so I told you there's the one folder I always connect to, is a broader lessons learned. The broader lessons learned document currently has over 200 lessons learned that are not project specific. So each and every one of the projects, as I mentioned, has a file that includes a lot of information, including the lessons learned that are just specific to that project. But in addition, anything that Claude and I, Chachapiti and I are learning as we're working, as far as the working environment, best practices, procedures, what tools to use for what, what tools not to use, how to avoid different pitfalls, and so on, gets automatically documented in that file without me having to do this.

10:45Again, the AI does it on its own. And then it knows how not to repeat these mistakes. It reads these three files before doing anything. So before doing any process on any project, on anything that I'm doing, it's going to read these three files. And then based on that, decide what to do next and knows how not to repeat mistakes and be focused on the next milestone that's going to provide us the most amount of value. The other thing, there is a tagging process. So this is the fourth component. The tagging process is a process that runs on every creation of every file. And also, I started with doing a retroactive pass to get all the data that wasn't there before.

11:27But basically, what it adds is it adds a short piece of information at the beginning of each file that tells the AI what's in it. So it currently has 15 document types that tells the AI what the file is, and then a short summary of each and every one of the files. The benefit of this is that the AI doesn't have to read all the files. It has to read only the first paragraph, if you want, like a menu of all the files, and then it knows which files it needs to read. It generates what's called a front matter, which sits in front of the document and tells the AI exactly what is in that information.

12:07And it's a very similar concept. Those of you who knows how skills work, it's the same concept. So the way skills work, there is a short one paragraph at the beginning of its skill that tells the AI what's in it. The same kind of thing, but for every document in the AI space that the AI is connected to. And then the last component is an index file. What is in the index file? It tells the AI basically everything that is in every document that it has access to. So the title, the type, the status, the date, the one line summary that I talked about before, the location of it, where it can find it.

12:42All of those is saved in one index file that is a simple MD file that Claude has to read every single time. In addition, there are unique specific index files for each and every one of my clients. because as I mentioned, these are saved separately and independent of the main registry, just to keep each and every one of them as a separate repository that doesn't touch anything else. And this is because of NDA purposes, and I don't want the information of one client to be reachable by a project that I'm doing for another client. So these are completely separate, but they have a similar concept that also maps all this information.

13:18Now, what other information is mapped in there. Additional sources of information exist beyond your hard drive. First of all, you have your shared drives, whether it's on Microsoft environment or in Google environment or another environment, you have those shared drives. Mapping those shared drives is also something that you can add in order to increase the access of AI to the information that exists over there. Again, you're not duplicating the files. The files stay wherever they are. The mapping of what are the files, where they reside, and what information exists in them can be mapped in your mapping.

13:53The next thing that can be a part of your knowledge base or knowledge graph is your recorded calls. I use Fathom, but there's many other tools that record and transcribe calls, which again means every meeting that I have, every single one of them in the last three years is mapped and tagged and the AI knows how to reach and get access to it. And then there is all your other systems, your ERP, your CRM, et cetera, each and every one of them in my universe, including my task management tool. I use ClickUp, but this could be Monday or Jira or Asana or whatever other task management tool you're using with your team to map different things.

14:27All of these can be mapped back in there. The reality, this can be very thin mapping. And the reason it could be very thin mapping is because all of these tools have great connectors, either MCPs or an API that you can connect through another source. So anything that have an MCP, I connect to the MCP. Anything that doesn't have an MCP, I have Claude build the API connector through NA10. And so I can reach those files and search them in real time and bring information. So you need a very thin layer of index, and then everything else happens through the MCP or the API. But basically every piece of information that tells the AI stuff about your universe and allows you to write and update information or capture and learn information in order to do a task can become a part of this knowledge base.

15:12Now, before I tell you how I build it, I want to show you quickly how can this work. And I want to show you one very simple example, just to prove my point. I asked Claude one simple question. What is the status of the website visitor pipeline, which is one of the projects I'm working on. I haven't touched it in a while, but you can see from a prompt engineering, this is a very shitty prompt. I'm not telling it exactly what I'm talking about. I'm not telling it which role it needs to take. I'm not telling it what I'm looking for, like nothing. It's a one line, not even a full line prompt. And then the response is found it.

15:44There's a dedicated website visitor pipeline project folder under the vault multiply. Let me look inside. Then it says consulted knowledge index that was generated on 2026.09.20 at 11 p.m. So that's late yesterday evening. Why? Because every single day, the index gets updated just before midnight to see any new files that were created. And then it read the Evergreen Website Visitor Pipeline and the Task Registry Website Visitor Pipeline. And then it told me everything that it found, where we stand, what actions I need to take, what is nice to know, and so on. All of this is available down to a very high level of detail, despite the fact I asked a very generic one line question.

16:28Again, this is very bad prompt engineering, but because it has great context engineering, harness engineering, and process engineering, it knows exactly what to do. And I can pick up any project, even if I haven't touched it in a year, from exactly the point that I stopped and Claude or ChatGPT or whatever tool I connect to it will know exactly what needs to be the next task, exactly what are the things it needs to connect to, exactly what information that happened in all the background for that particular client or process or project or anything that I'm working on and pick it up as if we just talked about it three seconds ago.

17:01So I know what you're thinking right now, like I want this as well. How do I build one? Great question. This is exactly what we're going to talk about right now. So I more or less told you all the different things, but I'm going to break it in order. The first thing you want to create is one folder that is split into areas. Those areas is something you need to decide and you need to think with yourself on how you want to break it up. There are several different consideration. The first one is data security. You don't want to keep sensitive client data with everything else. It's just a simple example, but there are many other reasons why you want to split data up.

17:33I'm running more than one business. So my businesses are split up. I'm running different aspects of my business. I have the consulting side. I have the training side. I have the workshops that I'm doing for organizations. I have the courses that are open to the public and I have the membership. So I have different components inside my business and different knowledge that is required for each and every one of them. And so it is built into different areas with some overlap between them that the AI knows how to share with the other components as well. So the first thing you need to do is to figure out what are these high-level categories or areas of knowledge that you want to split from one another.

18:10And you can do this together with the AI. It can help you get to the conclusion based on your current data. What is it that you want to structure in order to make these different areas of the world. I call them vaults in my universe, but what these five or different vaults needs to be. The second one is the instructions on how you want to work with Claude. There are two main places you can tell Claude, or by the way, any of the other tools, how to work with you. The first one is the general instructions of Claude that exist in the Claude settings. There used to be three different levels. One was for Claude in general, one was for Cloud Code, and one was for Cloud Cowork.

18:49They've merged it last week into just one. So that's the main set of instructions that Cloud will follow every single time. The other one is a cloud.md file that can be in any folder, and then Cloud will read that as well before doing anything in that folder. So if you tell Cloud to go to a specific folder, you will first and foremost look for a cloud.md file, and will read that. If it's a GPT, it's an agents.md file but it is the same exact trick again if you want to learn how to work with chatgpt and claude at the same time on the same information go and check out the previous episode that we have released so you need to think about what do you want claude to do every single time or chatgpt every single time you work with it one thing that we just talked about is reading and updating all the files that we mentioned that it needs to know how to do so once you do that claude will follow these instructions every single time.

19:45You will read the different documents you need to read, and you will update them as it needs to update them based on things that you are learning, that Claude is learning, or decisions that you're making, or progress that you're making, and so on. Now, if you're wondering what should be in the main instructions of the AI tool under the system instructions in the settings versus what needs to be in an MD file in the folder, it depends on how generic it is and how much you want it to do every single time versus is just for specific projects or topic. So if you want to do this for every single time, put it in the main instructions.

20:17If you want it just for a specific category, specific topic, specific project, then you can go ahead and put it in an MD file in that project. So this is number one. Number two is you want to create the files we just talked about. So you want to create a lessons file that is a generic one that sits on top of everything else that gets updated with things that matter for more than just one project. You want to create a evergreen document that has a built-in instructions and status for this specific project that you're working on. And again, I'm using the word project. Again, I'm mentioning that it might be confusing.

20:50It's not a project inside of ChatUPT or Claude. It is a folder that is a specific topic that I want to work on. Every one of those has one of those files and a task registry, which is a CSV file that has all the tasks that you're going to be working on. and it gets updated automatically. It doesn't need a million columns. You can decide what they are. I use a WBS structure. So basically one, 1.1, 1.1.1, 1.1.2, 1.1.3, and so on and so forth with what's the number, what's the description, what's the title, what's the status. You can add as many columns as you want for each and every one of those tasks.

21:29And then Claude takes care of everything else again or Chatubity or whichever tool that you're using. then you need to build a tagging process. You can do that together with AI. You can ask it, okay, here's what I'm trying to do. I'm trying to map all my files. Your file is going to be different than mine because you have different projects and different things you care about and so on. But if you have brainstormed with AI and anything, you can do the same thing here and just say, okay, this is what I'm trying to do. I want to map all the files in a way that will allow you, ChatGPT, Claude, et cetera, to find them easier.

22:01What do you think should be the different tags? What information we should put out there? And then once you agree what the structure is, have it backfill. So build a plan on all the files you already have, what and how to go and do it. Don't do it all at once because I don't know what's going to happen. You probably have hundreds and potentially thousands of these files. So you want to build a plan and then check the plan and check the first step, the second step after you get more and more comfortable that it's actually doing the right thing and you can test it and see that it's actually can find that information easy right now.

22:30Then you can go on large batches of information. Then once you have all of these things, you want to build the index and you want to build the index refresh script. So the index is really the map, the knowledge graph that sits on top of all of that. It's really that menu, if you want, or the table of context for everything that I have in my entire universe of knowledge, minus against the clients that have their own individual knowledge file of their own. because again, as I mentioned, I don't want data for one client being available when I'm working on a different client. So I have one for each and every one of my clients and then one general one for everything that is not client specific.

23:13And the AI knows how to read the right one and how to update the right one. And it knows how to update it on every single day running on a scheduled task. So every day or whatever hour you want, you can run it and it will look at what's new across your universe and it will update it into your knowledge graph. And now the AI will know how to read and find any piece of information it needs. It will know how to then find the relevant files, read the one opening paragraph, verify that that's really what it's looking for and open that information and do the same thing across data, across your universe.

23:46Like I said, go to your Google Drive, go to your OneDrive, go to your saved meetings, to your CRM, ERP, et cetera, and know how to pull and how to find information in there as well. Now, this sounds really simple, right? It's just a bunch of files. There's no databases. There are no APIs. There's no crazy sophisticated tools. And the reality is that was not my original plan. My original plan had a super-based database and a much more sophisticated process, including a graph retrieval engine that I was debating whether to continue and develop on my own or to use an existing one and customize it and server wiring that will allow all of this to work.

24:28And then after doing some additional research, I found out that it's actually not necessary, that it will be extremely valuable once I get to hundreds of thousands of files, which hopefully I will never get to. But right now, when I'm in the thousands, it is as efficient and as effective to just run locally on my computer with an MD file that is obviously backed up. So everything I told you, yes, it runs on local files, but everything is backed up to GitHub. In my case, you can back it up to OneDrive. You can back it up to whatever you want. Again, as long as you take care of those different universes separately and don't allow your customer data or your secret project data, whatever, to be in the same place as other aspects of your data.