The argument in one line.
You should disable auto-memory in LLMs and own a vendor-agnostic markdown folder structure instead, because it lets you switch between Claude, Gemini, Codex, or any other AI without lock-in while maintaining a persistent, interconnected personal knowledge system.
Read if. Skip if.
READ IF YOU ARE…
- You're using multiple LLMs (Claude, Gemini, Codex) across different interfaces and want a single folder structure that works with all of them without vendor lock-in.
- A knowledge worker managing 3+ concurrent projects who's frustrated with each LLM's proprietary memory system and wants explicit control over what your AI agents can access.
- You're running a small team or solo operation and want to spawn specialized AI agents (researcher, writer, hiring manager, orchestrator) from a single markdown folder without building custom integrations.
SKIP IF…
- You're a non-technical user who doesn't work in CLIs, GitHub, or markdown — this requires comfort with folder structures and agent configuration files.
- You're looking for a plug-and-play UI solution — this is a plaintext folder architecture that requires you to manually point different LLM interfaces at the same directory.
The full version, fast.
The video makes the case that the AI tool you use matters far less than the folder you point it at, and that letting Claude or ChatGPT manage its own auto memory cedes control you should keep. The mechanism is myPKA, a plain-markdown scaffold built around an orchestrator agent, a hiring agent, and a researcher agent that any CLI initializes itself into by reading a single adapter prompt file, with cross-linked journal, CRM, documents, and session-log folders replacing built-in memory. You switch off auto memory inside every LLM, route work through the orchestrator, close sessions to write durable logs, and stay free to swap Claude, Codex, Gemini, Obsidian, or a local model without losing context or rebuilding your knowledge base.
Free for members
Chat with this breakdown — free.
Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.
Create a free account →Where the time goes.
00:00 – 01:10
01 · Cold open + two promises
Turn off LLM auto-memory; the folder is agnostic — works with Claude, Codex, Gemini, or any local LLM.
01:10 – 02:36
02 · What myPKA is (and the GitHub drop)
Personal Knowledge Assistance, not PKM. Free download on GitHub; paid 20-lesson course at myicor.com goes deeper.
02:36 – 04:16
03 · Folder anatomy — Deliverables, PKM, Team
Walks the top-level folders: Deliverables (owner's inbox), PKM (knowledge), Team (agents).
04:16 – 06:00
04 · Meet Larry, Nolan, Pax — the 3-agent starter team
Larry = orchestrator (SPOC). Nolan = hiring agent. Pax = researcher. Nolan calls Pax to research new specialists.
06:00 – 08:00
05 · My Life concept + ICOR methodology
PKM/My Life folder = Goals, Habits, Key Elements, Projects, Topics. Load-bearing structure inherited from Paperless Movement's 4-year-old ICOR teaching.
08:00 – 10:10
06 · Journal flow + CRM auto-cross-linking
Hand Larry a screenshot or meeting note; Penn writes the journal entry, creates person + organization records, wiki-links everything automatically.
10:10 – 12:00
07 · Markdown vs SQLite + Obsidian as free viewer
Default is plain markdown. One prompt — 'switch to SQLite' — converts the vault to a database. Obsidian opens the markdown and renders the knowledge graph with zero config.
12:00 – 14:30
08 · Setup: duplicate folder 3x + the ADAPTER-PROMPT trick
Copies folder for Claude Cowork, Gemini, Codex tests. The ADAPTER-PROMPT.md file is what makes any LLM self-initialize — it tells each agent how to plug into the folder without custom skills or plugins.
14:30 – 17:30
09 · Demo 1: Claude Cowork init
'Initialize yourself inside this folder' — Claude reads ADAPTER-PROMPT, writes CLAUDE.md, becomes Larry. Then: 'Today I met Max Muster, works at Maximize.' Penn routes the journal entry; person + org get cross-linked.
17:30 – 19:30
10 · Obsidian knowledge graph reveal
Opens the same Cowork-modified folder in Obsidian. Knowledge graph already shows Dr. Schmidt → Clinic → Organization. Zero Obsidian-specific config needed; it just reads wiki-links.
19:30 – 21:30
11 · Team Knowledge = SOPs as the real skills
Team Knowledge folder holds SOPs, workstreams, guidelines, naming conventions. Tom's claim: this beats Claude Skills because SOPs are tool-agnostic and live in your repo.
21:30 – 23:00
12 · Demo 2: Claude Code terminal — 'who are you?'
Right-click → new terminal at folder → launch Claude → first prompt is literally 'who are you?' → 'I'm Larry, your team orchestrator.' Then 'do you know anything about Max?' returns the wiki-linked Max Muster + Maximize org from the previous Cowork session.
23:00 – 25:00
13 · Demo 3: Gemini CLI side-by-side
Launches Gemini CLI in a second copy of the folder. Same init prompt. Tom's stance: stop agonizing over which agent — any model with enough context is more than enough for personal knowledge.
25:00 – 27:00
14 · Session logs replace auto-memory
Tom shows /memory in Claude and switches auto-memory off. 'Close the session' triggers an explicit session-log markdown file in Team Knowledge/session-logs. Permanent, reviewable, portable.
27:00 – 29:00
15 · Demo 4: OpenAI Codex — already speaks AGENTS.md
Codex requires no init because it natively reads AGENTS.md. Tom drops a screenshot in, says 'create a journal entry about the launch of membership' — Codex routes to Penn, who writes the markdown file with the image referenced.
29:00 – 31:00
16 · Multi-agent same folder + Obsidian shows the result
Three different agent CLIs can point at one folder; they each create their own *.md config on top. Opens the Codex-modified folder in Obsidian — journal entry + embedded screenshot rendered cleanly.
31:00 – 33:00
17 · One-prompt SQLite migration + Gemini parity
'Switch to SQLite' — the scaffold's built-in SOP converts the markdown vault to a SQLite DB. Then shows Gemini's parallel result — same image-to-journal workflow worked identically.
33:00 – 34:33
18 · Pitch + close — course + GitHub + module roadmap
CTAs: download the free scaffold from GitHub, or take the 20-lesson course at myicor.com. Teases an AI Library of add-on modules (Slack integration, Telegram, etc.). Closes by reiterating: don't get stuck in vendor auto-memory, own the folder.
Lines worth screenshotting.
- Turning off LLM auto-memory and owning the folder yourself is the only way to maintain true portability across Claude, Gemini, Codex, and local models simultaneously.
- A vendor-agnostic plain-markdown folder structure with pre-loaded agent definitions (Larry, Nolan, Pax) works identically regardless of which AI CLI initializes it.
- Larry the orchestrator reads the agent index and routes tasks to the correct specialist — the same delegation logic a human manager uses, but implemented in a markdown file.
- Nolan the hiring agent spawns new specialists on demand by asking Pax to research the best available approach for the required role, making the team self-extending.
- SQLite as an optional upgrade from markdown gives the AI a more efficient structure to surface and cross-connect information — particularly valuable for large professional knowledge bases.
- The deliverables folder (owner's inbox) separates AI-generated outputs from the working folder structure, creating a natural review gate before any output becomes permanent.
- Obsidian can open the same markdown folder the AI agents use, making the knowledge graph visible and navigable without a separate import or sync process.
- A myPKA scaffold that compounds over nine months of conversation history gives the AI progressively richer context about the user's business without re-explaining on every session.
- An iOS app wrapping the same Supabase or SQLite data structure gives the founder mobile access to the same intelligence the desktop agents are reading.
- The PKA (Personal Knowledge Assistance) model replaces the PKM (Personal Knowledge Management) model by adding an AI action layer on top of the knowledge storage.
- Pre-loading agents with the ICO methodology's MyLife framework means the AI team understands the productivity system's logic before the user starts working, not after.
- Publishing the scaffold as a free GitHub download while offering a 20-lesson deep-dive course is the correct tiered distribution for a tool that is simple to install but deep to master.
Steal the format: free GitHub scaffold + paid course = trojan horse.
myPKA playbook
Ship the real, useful thing for free on GitHub — then sell the methodology that explains every folder.
- Open-source the scaffold (the WHAT). Charge for the 20-lesson course that teaches the WHY behind every file.
- Make the demo a stress test, not a pitch — Tom proves it works in 4 different agent CLIs back-to-back. The viewer convinces themselves.
- Use the ADAPTER-PROMPT.md trick in any agent-ready repo you ship: one markdown file that tells any LLM how to initialize itself. Portable across Claude/Gemini/Codex.
- Replace 'memory' with explicit on-disk session logs. Same idea works for JoeFlow's session-store — readable markdown beats opaque vendor memory.
- Adopt the SPOC (single person of contact) UX for multi-agent products. Joe's JACE/REESE/SAGE/RYDER setup needs one Larry-equivalent who routes — never make the user pick.
- Tom calls Claude Skills 'nonsense' — that's the contrarian wedge. Pick a fight with the dominant vendor pattern and you get airtime.
- Name your agents like a sitcom cast (Larry, Nolan, Pax, Penn). It makes a markdown folder feel like a team.
Terms worth knowing.
- PKA (Personal Knowledge Assistance)
- A reframe of the traditional PKM concept in which AI agents actively manage, cross-connect, and surface a user's knowledge rather than the user organizing it manually.
- PKM (Personal Knowledge Management)
- The practice of systematically capturing, organizing, and retrieving information for personal or professional use — tools like Obsidian and Notion are common PKM platforms.
- scaffold (AI folder)
- A pre-structured folder of markdown files that an AI agent reads on initialization to understand its role, the team roster, naming conventions, and workflow rules — making the setup portable and vendor-agnostic.
- adapter prompt
- A markdown file within the scaffold that instructs any AI agent how to initialize itself inside the folder structure — the single entry point that allows Claude, Gemini, Codex, or any other LLM to plug into the same system.
- orchestrator agent
- An AI agent that receives incoming tasks and routes them to the correct specialist sub-agent, rather than handling all tasks itself — the Larry agent in this system plays this role.
- auto memory
- A built-in feature in AI assistants like Claude or ChatGPT that automatically extracts and stores facts from conversations without the user controlling what gets saved or how it's structured.
- WikiLinks
- A double-bracket link format (e.g., [[Person Name]]) used in Obsidian and other markdown tools to create bidirectional connections between notes — enabling a knowledge graph of relationships.
- knowledge graph
- A visual map of how notes or entities are connected to each other within a knowledge base — Obsidian's graph view renders WikiLink relationships as a network of nodes and edges.
- SQLite
- A lightweight, file-based relational database format that stores structured data in a single .db file — used here as an upgrade from plain markdown when the AI needs faster, more efficient data access.
- single source of truth
- The principle that each piece of information is stored in exactly one canonical location, so all agents and tools reference that one file rather than maintaining duplicate or conflicting copies.
- session log
- A markdown file created at the end of an AI work session summarizing what was discussed, decided, and changed — used here as a persistent, structured alternative to auto memory.
- Gemini CLI
- Google's command-line interface for the Gemini AI model, allowing users to run Gemini against local files and folders from a terminal — equivalent in function to Claude Code or OpenAI Codex CLI.
- Codex (OpenAI)
- OpenAI's terminal-based agentic coding tool — separate from the original Codex model — that allows the GPT-4 family of models to read and write local files and run commands, similar to Claude Code.
Things they pointed at.
14:30toolClaude Cowork (Anthropic desktop)
21:20toolClaude Code (terminal)
23:00toolGemini CLI
27:00toolOpenAI Codex
17:30toolObsidian (graph view + markdown viewer)
Lines you could clip.
00:00
“Switch off the auto memory in your LLM, no matter if it is Claude or ChatGPT or whatever.”
22:02
“People worry about what agent should I use, what's the best model. If you're not into coding, I wouldn't worry at all.”
23:31
“You will never reach this efficiency level than having a proper local folder structure where you know how things are set up.”
28:34
“If you want to have proper databases, it's just one prompt to switch to SQLite.”
23:15
“The auto memory will just randomly grasp the things that AI considers that is useful to you.”
Word for word.
Read-along
Don't just watch it. Burn it in.
See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.
metaphoranalogy
00:00In this video, I want to talk about two things. First of all, to switch off the auto memory in your LLM, matter if it is Claude or JetGPT or whatever. There's a much better way to control the persistent memory of your AI.
00:13And the second one is that even that I shared a lot about Claude in the previous videos using my local folder, we don't need to use Claude. That's why we built this folder structure from the ground up independently from any LLM.
00:28And in this video, I show you proof that no matter if you're using Claude, Codex, or Gemini, or any other local LLM if you might need it, as some have some security worries, they rather want to use a local LLM. All of this works perfectly with this folder structure that I'm showing you in this video. In the end of the video, you get a complete different perspective on how you use your AI with your professional work to get things done and manage your knowledge.
00:55It replaces your personal knowledge management. As we say, it's now p k a, . And that's what we actually also show in the new course.
01:04Our members can directly download the folder or you go to GitHub and you can download it from complete free. But if you really want to dive deep, this is what the course is here. And each of the 20 lessons explains step by step how this folder is set up and what we add to the different files.
01:21So if you move forward to this, you see here there's then step by step the folder structure building up, and you have the information what's building, and we explain step by step in much more detail than in this video is possible because it takes more time. How everything works under the hood and how this folder is built.
01:39So here is the folder that we are talking about, and I will show you step by step what is in the different files and how your AI will work this folder. And in order to better present this to you, I will use our course that actually in each lesson shows you the different files and folders that we add over time to this, as it is also better when I click into any of these to show you the actual content.
02:01So the folder that we are building to have everything local and AI independent is this. You have this deliverables folder where whenever your AI agent is working with you and the team delivers something, it goes into this folder.
02:15It's the owner's inbox that we have here. So whatever the team brings back or gives you to review, it ends up in this folder. Then we have the team folder, which holds the different team members of the team.
02:27I made several videos about this already. In short, the most important three agents in any of these folders are Larry, the orchestrator, who is just whatever you give it to him, he will look into the agent index and sees what agents are available to work with and what he should hand over to this agent.
02:47Then we have Nolan. This is the agent who actually hires new team members if necessary.
02:54So let's say I need to I I want to build an app. We need might need a front end developer. This is not their out of the box.
03:01This is something that Larry would recognize that we don't have an expert in this field of expertise. He will reach out to Nolan and ask to hire somebody, And Pax, the third agent in this constellation, is the researcher.
03:15So Nolan will ask Pax to go online and research about the best front end developer to build this, or the best journal writer, or the best designer, or the best accountant, or whatever you need for your personal needs, these three agents alone, Larry, Nolan, and Pax, will be enough in this folder to start building your AI team in a way that you need it.
03:38Just to give an example how this looks like when you work with them, here's just a set of different agents that I'm personally using that we also provide to our members already to download, but I always recommend just build them for yourself because using the team as I showed with Larry, Nolan, and Pax, it's much more efficient to build it custom made for your specific needs.
04:00Yet, if you want, you can have a starting point from my established agents that I'm already using. So you see in this final folder that we end up in the course, there's also Pen, a journal writer, and Felix, a front end developer. And these are just examples that are created during course, and this is a personal knowledge assistance.
04:20So this means we have a PKM system inside this folder that is very sophisticated because it is based on our my life concept, which is part of our ICO methodology.
04:32And this is why this goes way beyond just giving you a empty folder and say, okay, you have some goals, habits. No. The team actually knows how to manage this my life.
04:42If we go into the PKM Leica Pro course, there's the MyLife session, and here's where we explain the details about how this works, and how you apply this literally to any tool.
04:52This is now an example with AI, but we've been teaching this already for over four years to our members, how you can become more productive, how you manage your personal knowledge just with this part.
05:04But you see we are going way beyond. We have note taking, like a pro task management, like a pro project management, like a pro, and then automation, which includes the AI layer two.
05:14So it's literally a productivity system end to end for any business size, entrepreneurs, and big corporates alike. And this is literally what this team in this folder is capable of to manage.
05:24It will be the digital note taking and the PKM area. So what does this mean in action? Whatever I hand over to Larry with the Spock, , whenever you open your Claude co work or your your Claude in terminal or your Gemini or whatever, and I show you in this video later on how this actually works in action.
05:45I just want to give you a bit more theory upfront before we just dive into this. Larry will recognize what you're handing over there, and then he also is aware of all his my life structure. So let's say you have a screenshot about something that's inspired you, and that's not something that affects your life yet, but it's something you're interested.
06:03Larry will decide that this is becoming part of topics. So he will create a new topic in there, and he will start managing all this. This new folder structure, by the way, is based on markdown only.
06:15So this means, and we can do this in this video later too, we can perfectly open this up in Obsidian, and Obsidian understands the connections of everything, and you can use it with Obsidian if you want. And yet, there is an additional lesson already included in the Scaffold two. We can ask Larry to upgrade the whole folder into an SQLite database structure, which is much more efficient than using something like Obsidian when it comes to professional knowledge management and make things much more efficient to access for the AI.
06:45Don't worry if SQLite and database seems complex to you. All this does, it gives a much easier structure to access for your LLM to be much more efficient to surface information, to cross connect information. And yet it perfectly works fine for personal knowledge, also just using markdown.
07:02That's why this new folder structure that we provide here perfectly works with Obsidian, and you have access to all the file, you control everything that you have. Then we have here this journal, and you see already there's the first example entries where we have something like this.
07:15And how this works is we can just hand in a screenshot, as I mentioned, and Larry together with Penn, who is the journal writer, will create a journal entry out of it. But this goes way beyond just writing down what you have done this day. If you mention any person that is relevant, if you hand over meeting minutes or something like this, the team also has a CRM, which manages your people that you have contact with and the organizations, and all of this is cross connected.
07:45We will look at this in the knowledge graph in Obsidian in a moment. There you see the visual connection, but this is just it, and doctor Schmidt is actually, uh, one of the examples that our members go through in this course.
07:58Because in this lesson, see, there is this where this is gets created just by simply this mention here. It will then literally create a doctor Schmidt, which is the details about the person connected to the clinic this person is working. So you see their Wiki links.
08:15Everything is cross connected, and this is what you will do on autopilot. So if I have a business card from somebody, I will just hand it over.
08:22It will create all this automatically. And whenever you do daily journaling, it starts then to extend your people database and starts cross connecting everything.
08:32But this is why it's crucial to have the my life structure in here because this represents your full life end to end. You will have always interests, which is the topics.
08:42You have key elements like family, your business, your day job, things like this. You have goals that you set for yourself, and you achieve these goals by establishing habits.
08:52Everything, all of this is just information. This is not task management or project management. We say here there are project habits that sounds like action, but in the end, it's the information about these action items, and yet you need to organize these actions as we teach it in code two with tools like Todoist or with ClickUp because this is a complete different complexity level how we represent action in our lives.
09:21That is much more efficient doing it myself than letting AI do it. But what AI is really powerful in is understanding cross connections of your knowledge of the things that you gathered over time.
09:31Then there is a documents folder. So if I scan any documents in my case, they end up here. So it will just get organized inside documents.
09:40I will have to have things there. Images, screenshots, anything like this will be stored there. You might just know this from Obsidian.
09:47In the end, we are just referencing these things and can embed this in our notes, but this is a perfectly structured organization, and the same would work if you use SQLite databases.
09:58It would just reference as the original images, so AI will be always able to surface any PDF document that you're looking for, anything that you had, because it has everything indexed, and these are these index dot m d files. So now in the beginning of the video, I mentioned it's completely AI agnostic. I just have a local folder that I already opened here.
10:19What I want to do now, I show you the independence from any AI tool. So I will copy this in three times, and I will show you the content is always the same.
10:30I just changed the number in the end to differentiate between those. But you see that the content is always the same. And I will show you that this perfectly works no matter if I use Claude, Codex, or Gemini.
10:42Any LLM that can access a local folder, even Perplexity Computer would be able to do this, can work within this file structure. And I even show you with Claude Cowork to get started that this is possible.
10:55Let's make an example file just for Claude Cowork. Here we are. Yes.
11:00That's the one that we will use with Claude Cowork. I go into Claude Cowork, choose a different folder, go to desktop, and here's the one that we will use for Claude Cowork. Now I loaded this in.
11:11I'm working in this folder now. See? I even use Sonnet here.
11:14I can use Opus, but I just want to show you that really with the basic mail. I saw it many times. I'm I'm on a pro subscription for Claude.
11:22Of course, you are using the max plan, and this works fine. No. This is in fact saving you a lot of tokens because everything is set up very efficiently in this folder, and by the handover to the different agents, you are saving a lot of tokens compared to always letting it scan through everything from the ground up.
11:40We go to this activate your scaffold. How does this work under the hood? Well, this is where this adapter prompt dot m d file comes into play.
11:49That's in this folder. So if you look into this folder, see here, this is the adapter prompt dot m d file, and this is what is in there. And what it does, it just tells any LLM to initiate itself inside this folder, but keep things very simplistic and rather forward and reference these agent dot m d files instead of creating custom skills.
12:12So you might have seen endless videos where you can download skills, you can download plugins, and whatever for Claude.
12:19And this is nonsense in my opinion because those are all premade settings for your Claude that might be overkill, that is loading a lot more tokens in that you actually need. And that's why we keep this very slim, and it keeps us independent so we can switch LLMs models from Claude, Gemini, and so on by simply reading out this adapter prompt.
12:43So I can just say, initialize yourself inside this folder. I think this should be more than enough testing it for the first time, but I'm pretty sure it will get it.
12:53So here, it's actually looking now for the claw dot m d, so it creates this now. There was no claw dot m d file in the first place. It runs now the init skill.
13:02So in claw dot code, you would says say slash init. So what it's in the back end actually is doing now, and we are here in Claude Cowork, you see, it's still working. It looks into the folder structure and starts to understand what is going on there.
13:15It will find this adapter prompt, and now will understand how it can plug itself into this local folder. There we go.
13:23See, it found the adapter prompt template, and this is this defined everything for it. It created the claw dot m d file.
13:30Here you see what's going on. It's creating it right now, and there we go. It created this new claw dot m d file in this folder.
13:37And if I open this up, you see now it just analyzed what is going on. It recognizes he's Larry now, that he is now looking into agents m d and not claud on d files and so on. This is all what happened by just saying initialize yourself in this folder, and you can do this with any folder because this is setting your claud up in a way so it knows what is the content of photo.
14:01We're doing nothing different here, but you see already he understands now some basic rules of iCore, single source of truth, where to find things, where to store things. All these things are now there. And if I now say, who are you?
14:14He already initialized himself. He now is Larry, and he's the team orchestrator at his PKA.
14:20So and I'm running Claude with pen, pegs, Nolan, ready to go. These are the basic agents available in the scaffold, and now I can say, today, I met Max Muster, and he's working in a huge corporation called Maximize.
14:35We had a good conversation about some investments that we could do into one of his businesses. Alright.
14:41Something random, and let's see what he will just react with in Claude Coburg by simply giving some generalistic things.
14:50So what I'm expecting now is that he's routing it to the right people, and yes, you can see he's routing it to Penn. Penn is the agent doing the journal entries as you might you remember. So he perfectly hands it over to the right agent.
15:04And remember, we don't have any specific agents in a dot cloth m d file or anything complex because everything is based on custom folders that we have full ownership and control of, and now I can could simply switch over from co work to something else.
15:21We will see in a moment. Now they go. Everything is cross linked.
15:24He created this, and now let's open this up inside Obsidian because this shows you now how flexible this folder structure actually is. So this, by the way, is just an example of my personal knowledge assistance that I had already in place, and I perfectly was able to import all the information, look at this, that I personally have.
15:46So you see, I'm using this on a daily basis. Paco is using this on a daily basis, not inside Obsidian. We don't need it.
15:52We have our custom built apps around it, but you could per perkly use Obsidian if this is something that you prefer and you like these knowledge graphs and so on. And here you go. There you see there's this index file.
16:04That's what people like to see. Even if you're importing these things, you see this grow all the fancy GIF animations, but in the end, there's no magic behind. In fact, Parkour and IB use SQLite other bases and have a much deeper interconnection of all the different entities than we could ever have in Obsidian.
16:21This being said, let's go here. Open vault. This brings up this window.
16:26I can click open. I go to desktop, and here's this folder that we created. When we go in here just to show you, there's this PKM, there's the CRM, there's people, and there is now Max Musta.
16:37So it's the correct folder. I just opened this up in Obsidian, and you see the full folder loads in. Now we are here.
16:43CRM, people, Max Musta. I can open up.
16:46You see it created already these properties. So I could now just talk to my AI and say, well, I got the email or phone number, whatever, and I can start storing this like I would manually do this inside Obsidian, but you see it's already cross connected because here is the journal entry.
17:04So I can click, and here we are, journal 2,026, and I'm in my journal entry, and here's the cross connection. And if we go here now to this knowledge graph, you see things are already interconnected.
17:16So you saw, I started with Cowork in this folder. I didn't do anything with Obsidian to this point, and now I just open it up as a new fold. And as it is all using WikiLeaks, everything is interconnected out of the box.
17:28So here you see how things are connected, the different agents, agents index, that just visualizes the context in this folder. But here we go. There's doctor Schmidt, and he's connected with the clinic, which is the company.
17:42If I click here, you see that's the organizations, and this is stored as an organization. And here we go.
17:47There's Max Musta, and there's the interconnection between Maximize, the company that we mentioned. And here's the entry that I can go to and you see the full context there.
17:56So you can imagine now handing over just your meeting minutes, the transcripts that you from a meeting to your AI, and it starts now on autopilot, interconnecting all these things.
18:08And that's the powerful thing that we have here and that you can see also in here. But this is something that happens on autopilot, and obviously this goes way beyond just this because I showed you in another video that I built the whole membership platform and the courses and all this using just Claude Code, but it was not just using the chatbot of Claude Code.
18:28It was literally this setup that you have here, where inside the team knowledge, there is where standard operating procedures that describe perfectly what the different agents should do, which you might know as skills in Claude, but this is the real world equivalent.
18:44So if I'm working with human teams, I would create standard operating procedures to tell my experts how they should work inside the company. And now having it separated in a team knowledge folder, this makes it much more efficient because I built a that I just need to update this one file, and all the other team members will leverage from this if they need to use the same standard operating procedure, for example.
19:08And then we have also work streams already integrated here. One work stream is already the daily journaling, and this is describing the process map how the agents work.
19:19So do you need to create these files? No. Because this is what Larry will do.
19:24And the combination of Nolan, who is hiring the agents, and creates this skill set for these agents. So Nolan, the hiring agent, has in his his agents MD all the instructions how to hire a person and the references in here. So we have also guidelines here.
19:40Okay? This is the naming convention, for example, that many more different agents are referring to, not only one. And that makes it so efficient to set up because then just using an index file, they know very quickly where to look into, and that saves so much tokens.
19:56So just with this in mind, I have here now a team inbox where I can hand over, for example, my scanned documents. I can even directly scan in here because remember, we are here on a local folder on my drive.
20:09I can synchronize this with Dropbox or a g drive or whatever, have it backed up and accessible from anywhere else. I saw people saying that they have issues using something like Dropbox or iCloud that things are breaking. This only happens if you try to use the same folder on different devices in parallel.
20:27But if I work on one device and then later on on a different device, this is never an issue. So I can perfectly work this way. And now this is still the one folder.
20:36This is now set up with Claude Cowork, but I can perfectly now right click on this folder. I say new terminal at folder, and now I can use the terminal version of Claude. I just launched Claude, and here you go.
20:49I just say, who are you? Without doing anything else, the first prompt is who are you? And he immediately says, I'm Larry, your team orchestrator, and so on.
20:58And now I can say, do you know anything about Max? And here we go. He found Max Musta, all the information.
21:05He found the WikiLeaks, how it's interconnected. He recognizes that there are no details yet, and that's where I prefer to build now a custom interface to represent this databases or whatever you want, and that's something we can do in a different video.
21:19And in fact, that's something we will provide as add ons inside our membership too, where you can simply plug things in, like Slack integration. Okay? So I could talk to this local folder through Slack or, uh, the Telegram integration that so many people seem like to use.
21:34Also, I would prefer Slack in this case because I can have much more structured conversations with my AI. Well, a whole other video about this. So now we are here with Claude Code.
21:44We, in the beginning, created these other copies of the same folder. So let's see what happens if I launch this inside inside Gemini, and I'm opening here the CLI, don't worry if this is all too complex.
21:58Uh, all I'm doing here is launching Gemini inside CLI. I could also go to antigravity, opening it up there.
22:06Whatever can access this folder, it will work. So I'm here now in Gemini. He's now inside this folder, and now I can just say in it, because there's a command in it, or I would say, like in Cowork, initialize yourself inside this folder, and he will do exactly the same thing.
22:22That's the crazy thing. People worry about what agent should I use, what's the best model, and so on. If you are not into coding and you really want to get on the edge of the best thing, I wouldn't worry at all.
22:34Because most of these models are more than enough to handle your personal knowledge, especially if you have it set up the way as it is in this folder because everything is interconnected and it makes any model very efficient.
22:47So while Gemini is launching here on this folder, let's open up in parallel another terminal on codex, and codex is in fact JetGPT. Alright? That's OpenAI, and codex is just a version like we know from Gemini and Claude that allows you to access this.
23:04JetGPT is just a chatbot, something we never use because getting back to the beginning of the video where I mentioned you should switch off the memory. This is where you have now the memory in the local folder. So if I go here again into the terminal, and I launch Claude again in here, you can have slash memory.
23:21And here you can switch on or off these memories. See? Then I like to have this switched off here.
23:27Why? Because the auto memory will just that is useful to you, And especially if you do this now in Claude, and with all the things that they are publishing about the custom agents, and you can give it names and so on, now you are stuck inside Claude.
23:45Just keep this in mind. The more you use this desktop application, and you now use the convenience that they provide to you here, and use this auto memory and so on, you get a feeling of, yeah, it improves over time, it remembers me and so on, but you will never reach this efficiency level than having a proper local folder structure where you know how things are set up and make AI work in the way that it works best for you.
24:12And that's why I always switch off any auto memory that these AI agents provide, because what we have instead is just close the session, and we can have even the commands built therein. And what it will do, whatever you talked with it, because we didn't talk much to in this, we just launched memory in this session here, But it will now go through the directions.
24:33It will understand what you discussed about, and it will make a session log, and you see here, it's already finding everything perfectly. So if we can look into again into the folder structure, and we go into the team knowledge, there's the session logs, and here you see the logs, and there's nothing in there yet.
24:49He's creating it now, and this ensures that this is proper permanent memory. Okay?
24:55So you see it created now this MD file. I can perfectly open this up and see what was the conclusion of the team after the session. And this becomes really powerful.
25:04I'm talking here about just starting out with the folder, but imagine that you use this for weeks. This builds up over time. Anything that you say, oh, you went off rail, optimize this.
25:14Whenever you say then close the session, it will create these logs or just mid session, you say, keep this in mind, it will create these log files and therefore can go backwards in time when you say, well, last week things worked much better. It can review now all these session logs to understand what was going on.
25:31And this is much better than the auto memory that is just random and never so comprehensive and context connected than building it this way. Now you see for the other two agents they launched, so we have here OpenAI Codex.
25:46I can also say in it, but I also say the same as I said before, initialize yourself inside the folder. So this one is in this folder and here's Gemini in this folder and you see here, it read all the files, it understood this, it found the adapter prompt automatically. You see, I didn't do anything.
26:03I didn't give it a specific prompt or anything. It just looks through the folder because AI is intelligent enough and has enough context window to understand whatever is going on in this folder. And now it launches itself in there, and I say, yeah, just do everything everything necessary, and it will create a Gemini MD file now, very likely.
26:23So if we now open up this folder from Gemini, we see it created now a Gemini dot MD file. For anyone who has now the questions, can could I use several different agents in the same folder?
26:34Absolutely, you can. I could now use just one folder, point all the three of them onto the same folder, and I say initialize yourself, and they will just create what they need in on top of what is already there. But the beauty is everything links in these basic folders now instead of loading everything into these claw dot m d files and so on.
26:53It's offloaded into external information and it's cross connected, and that makes it just a starting point for these LLMs to initialize themselves to understand what's going on in the folder. And then they work perfectly inside this folder structure without the need to have a see, there is no in this one, for example, there's no dot clause file, and yet it recognizes that there are agents.
27:15It should work, and it should launch the different team members. But there's no dot Claude MD file or created any skills or agents because this lives all in here, in these agent MD files, and they get loaded the moment Larry needs support from these team members.
27:31And here we are, Codex. I said initially as yourself, he already recognizes all of the box that he is Larry because he read the files. He loaded the routine table.
27:40And the thing is, Codex is actually using agents.md, so there's not even confusion going on. It just uses already agents.md, so that's why there is no big initialization necessary.
27:52And now I can just use a random image screenshot. I just paste it in here.
27:56See it adds as an image and I say create a journal entry about the launch of membership.
28:04And now it uses this image. Here we encodecs. Keep in mind.
28:07And I can do the same here in Gemini, so we have it all. So I again use this image, just paste it in here, and I say the same thing.
28:16I copy paste what we have in here. Just paste it in here and do the same. And here we're in Gemini.
28:21Here we're in Codex, and it will perfectly work. I'm routing this to Pen. See, it recognizes the team roster, who is responsible for what.
28:29This is a capture request with a screenshot. Pen will create today's journal. Pen has enough.
28:34They read a look at the image, what is there. He creates now the journal entry, and he created it. He gives me now the link to the file.
28:41So I open up this folder in the PKM journal. So if we go here, PKM journal February.
28:48Here we go. There's the launched based on my name conventions. If I open this up, there we go.
28:54There's all the information that we have. It cross connected the wiki entries.
29:00Okay. But it didn't save the image yet. That's something you can perfectly do.
29:04It just didn't have the access to this image. So how I would usually do this, I can always open up this folder, have the team inbox here, and drop the file in here, and then I say, I've provided the image in the team inbox. So in this context, that's how I would start working on it.
29:21Now it has the the image available inside the team inbox. I could hand it over in the team inbox and say, create a journal out of this. But also, if you're using this, let's say, in a with computer access, it would have perfectly access to the image and, you know, integrate it.
29:37So you see it disappeared already from the team inbox, and now it moved it into the right position. Here we and then let's see in Obsidian. Here we are.
29:45This is the folder that we the codex was just working in. Here's the entry, and here's the screenshot with the context below. So if you want to have an interface out of the box, this new scaffold perfectly works.
29:57You saw I didn't use any Obsidian CLI, anything. It's not necessary.
30:02Obsidian is just recognizing WikiLeaks and even the properties it's recognizing because it's set up based on Wiki handling.
30:10There's nothing specific to Obsidian to understand this, that that codex is now trained to use Obsidian. No. You can now simply either open this up in Obsidian and visualize it or use any other tool that can read markdown or Wiki files, or build your own interface.
30:25It's really up to you, and you're really independent. And if you want to have proper databases, .
30:33So I could say now switch to SQLite. It will know what this means because the instructions are also in the and now it will convert this markdown structure into SQLite if you need it.
30:45For most people, the markdown is more than enough, and they prefer rather to see it transparent like in here, rather than having a database, and then you have need different visualizations. And for Gemini, you see, the Gemini was able it it used the image. It moved it into the right folder, and it's already there.
31:03So this was the Gemini folder. Let's also open this up in Obsidian. Here's now the Gemini folder that worked in here.
31:08So you see here is the Gemini MD. We are in this folder. You can perfectly see what's going on.
31:13And if now go PKM images, here's the image saved.
31:18And with Gemini, you have exactly the same solution. So now you can imagine for yourself if you start working with this all use cases. All you really need is to hand over whatever you have.
31:28Daily journaling was never that easy. I can talk to AI. I can audio capture myself.
31:34It's on the go. Many things that I can do now this way, and I can endlessly expand my folder here.
31:41But in the end, I don't care if one of these companies decide to increase token pricing or anything like this. I can keep switching to the cheapest if it is necessary or depending on security or having a local LLM. So many things.
31:56But as you have seen just with these three examples here, it always works the same out of the box, and it's all based on local folders and instructions that are based on a tool agnostic methodology that defines this work between the agents end to end, but also how to structure your PKM based on the MyLife concepts and all the things that we teach in my iQO to our members and here on YouTube, obviously.
32:23If you really want to dive deep into the how this was set up, you can go into our new course that's now available to our members where we show step by step how this folder structure was built. So in each lesson, you see how these different folders get added and why we add these and what is in these folders and why are there because there are descriptions below that dive really deep into this.
32:46Or you simply go there and download the Scaffold and get started right out of the box. And then there is this AI library where we also will add new modules that allow you now to expand the Scaffold with like a with a block system. So you can say, want have this agent.
33:02I want to have connection to Slack. I want to have x. That's where we will have a lot more modules coming up where we first explain how it is done, but also provide you just an out of the box folder extension or prompt, whatever you need in order to get this started.
33:17But I hope even without going through the course that this video gave you really the the right inspiration how you can build this up. Obviously, we are using the my life concept and that that was not the topic of this video today, how everything gets organized based on goals, habits, the key elements, projects, topics, but how the team works, how to make a folder independent from your AI that you're using.
33:42This was the key message that I wanted to get across in this video, but also why you shouldn't use these auto memories and things like that. And maybe think again if it might be worth for you to also switch to the terminal mode, even that it is doesn't look so fancy, you at least have full control in a raw format to your folder structure.
34:02Or you integrate it into Obsidian as I showed you in another video, how you can perfectly run your AI inside Obsidian, inside the terminal there, and access the folder structure there. This is the follow-up video as well for all these people who wanted to see the team working inside Obsidian.
34:19Well, and let me know in the comments below what you think about it. Are you already using the Scaffold? What's your experience with it?
34:25And I cannot wait to catch you up in another one where we will dive deeper into any of the extension modules that we have available. Catch you up in the next one.
The Hook
The bait, then the rug-pull.
Tom Solid opens with a two-headed promise — switch off auto-memory, and stop picking sides in the agent wars — then spends 34 minutes proving the same plain-markdown folder works identically in Claude Cowork, Claude Code, Gemini CLI, OpenAI Codex, and Obsidian. The free GitHub repo is the entire thing; the course just teaches the why.
Named ideas worth stealing.
01:00concept
PKA — Personal Knowledge Assistance
Tom's rebrand of PKM: instead of you organizing notes manually, the agent team does the cross-linking on autopilot.
Steal forany rename play where you reposition an existing category under a new acronym you own
03:20model
Larry / Nolan / Pax / Penn — the 4-agent starter team
- Larry — Orchestrator (SPOC, single person of contact, reads agent-index)
- Nolan — HR / hiring agent (recruits new specialists by writing new agents.md files)
- Pax — Researcher (Nolan's research arm)
- Penn — Journal writer (handles screenshots, meeting notes, cross-links CRM)
Named, role-based agents stored as markdown files. Each one is just an agents.md with a role description and SOPs.
Steal forJACE/REESE/SAGE/RYDER — Joe already does this; Tom's hierarchy (orchestrator → hiring → research → specialist) is the playbook
06:20model
My Life concept
- Goals
- Habits
- Key Elements
- Projects
- Topics
The PKM/My Life folder schema. Inherited from Paperless Movement's ICOR methodology taught for 4+ years.
Steal forPKM scaffolds need a top-level schema before agents can think — copy this 5-folder model or invent your own
04:08concept
SPOC — Single Person Of Contact
Larry is the only agent you talk to. He routes to specialists internally. The user never picks which agent to invoke.
Steal forany multi-agent product UX — one chat interface, not an agent picker
20:07concept
SSOT — Single Source Of Truth
Team Knowledge holds SOPs once; every agent references them. Update one file, all agents adopt the change.
Steal forops docs — kill the duplicated knowledge across per-agent system prompts
13:40concept
ADAPTER-PROMPT.md trick
One markdown file in the folder root that tells any LLM how to initialize itself. The folder is the WHAT; this file is the HOW.
Steal forshipping any agent-ready repo on GitHub — one init prompt portable across Claude/Gemini/Codex
25:00concept
Session logs > auto-memory
Explicit, human-readable markdown logs of each session in Team Knowledge/session-logs. Replaces opaque vendor auto-memory.
Steal forJoeFlow's session-store work — surface session summaries as readable files, not opaque blobs
How they asked for the click.
VERBAL ASK
33:00product
“If you really want to dive deep into how this was set up, you can go into our new course that's now available to our members where we show step by step how this folder structure was built. Or you simply go there and download the Scaffold and get started right out of the box.”
Soft. Free GitHub download is the headline; paid course is positioned as the 'go deeper' option for people who already trust him. Two CTAs threaded through the whole video, hard pitch only in final 90 seconds.
MENTIONED ON CAMERA
FROM THE DESCRIPTION
PRIMARY CTAWhere the creator wants you to go next.
OTHER LINKSAlso linked in the description.
- 👉 GET THE FREE FOLDER ↗
- 🎙️ Voice Dictation – Wispr Flow ↗
- 🧠 PKM Deep Thinking – Heptabase ↗
- 🌊 PKM Shallow Thinking – Tana ↗
- ✅ Task Management – Todoist ↗
- ⏰ Time Management – Sunsama ↗
- 🚀 Project Management – ClickUp ↗
- 👥 Team Deep Thinking – Miro ↗
- 📧 Email Management – Superhuman ↗
- 🤖 AI Meeting Notes – MeetGeek ↗
- 💻 MacOS Productivity Boost – Raycast ↗
- 🎥 Our primary video editing software is Descript ↗
- The AI Team Setup Nobody Talks About ↗
- My Claude AI Setup Does What a Human Team Would ↗
- Claude's Global Memory Is Fighting Your CLAUDE.md ↗
- Obsidian With Claude (The Setup I Said You Didn't Need) ↗
- Productivity for Professionals (Systems building Channel) ↗
- Think. Map. Execute. (Tools Channel) ↗
- Productivity like a Pro (Podcast Channel) ↗
- Tom at Paperless Movement (This Channel) ↗
- Linkedin ↗
- Linkedin ↗
Visual structure at a glance.
