38:55Systems Made Better · Tutorial
How I Cloned My Favourite Experts Into AI Agents
A 39-minute live build of a knowledge-grounded Notion AI specialist, using Seth Godin as the source material.
June 5thThe argument in one line.
An AI agent that never gets lost doesn't need better memory from OpenAI or Anthropic — it needs a router map written in plain files and folders that any LLM can read and follow.
Read if. Skip if.
READ IF YOU ARE…
- You've been frustrated that AI forgets context between sessions and wish you didn't have to re-explain your projects every time.
- You're using multiple AI tools — Claude, Codex, Cursor, local LLMs — and want your context to follow you across all of them.
- You build solo and want a system that handles task routing, project navigation, and handoffs without paying for an expensive memory subscription.
- You've seen AGENTS.md or CLAUDE.md files and want to understand the deeper organizational philosophy behind them.
SKIP IF…
- You want a polished, structured tutorial — this is an enthusiastic live brain-dump, not a stepwise course.
- You're already deeply familiar with Jake Van Clief's system and building sophisticated agentic folder structures.
The full version, fast.
AI agent memory isn't a technology problem — it's an organization problem you can solve with plain files. By placing AGENTS.md router files in every folder, you give any LLM a map that tells it where to go, what to do there, and what to do if it's lost. The host live-demos the system across four different AI apps and shows that switching between them requires no re-onboarding. The core workflow uses two keywords — 'pickup' for a start-of-session context brief and 'handoff' for an end-of-session state save — both triggered by single words that hit pre-written file-based instructions rather than manual prompting.
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 – 01:36
01 · Why this AI system matters
Personal backstory of 8 months of vibe coding frustration. Strong opening claim: this is the most valuable AI video on YouTube.
01:36 – 05:45
02 · File-routing experiment
Live demo of three nested AGENTS.md files routing an agent through a skill file to write 'control your agent' one word at a time.
05:45 – 08:05
03 · Testing with Claude
Runs the experiment in Claude Code. Agent completes the multi-hop chain unprompted. Done, boss.
08:05 – 10:09
04 · Testing with Codex
Same experiment in Codex. Emphasizes agents can execute goal-level tasks without being told the goal at runtime.
10:09 – 13:04
05 · What an AI agent really is
Demystifies agents — it's just a prompt. Shows that an AGENTS.md file with the same prompt as a ChatGPT GPT produces identical behavior.
13:04 – 15:44
06 · OpenClaw and AGENTS.md files
Opens OpenClaw's pre-installed AGENTS.md to show professional AI products use the same pattern. 'If you can read these files, you can build your own.'
15:44 – 17:04
07 · My personal folder system
Live look at the host's actual iScaleLabs folder structure. Each folder has its own AGENTS.md with sub-instructions.
17:04 – 20:12
08 · The map: keeping AI from getting lost
Deep dive on the map layer. Building-floor analogy. ROUTER.md becomes the master org directory.
20:12 – 22:05
09 · Routing to the right project
Says 'I want to work on iScale merch' in the lobby folder. Agent navigates to the correct project folder without being told where it lives.
22:05 – 25:05
10 · Tasks, automations, and org routing
Throws off-topic requests at the agent — task list, automations. Agent uses router to find the right file every time without prompting.
25:05 – 30:07
11 · Root law, gates, and Router.md
Full ROUTER.md map revealed. Gate logic explained: if task isn't in this folder's scope, return to root. Routing is self-correcting.
30:07 – 34:32
12 · Pickup and handoff workflows
Two-keyword workflow: 'pickup' = read-only context brief at session start; 'handoff' = full state save at end. Both triggered by one word.
34:32 – 37:08
13 · Switching projects without losing context
Mid-session project switch from iScale Merch to PinTwist. Handoff writes context. Later pickup in PinTwist recovers 'PinTwist World, VR walkthrough' without re-prompting.
37:08 – 44:02
14 · Using the system across different AI apps
Full cross-app demo: VS Code, Cursor, Hermes. Same AGENTS.md pasted into each. All recover PinTwist World context.
44:02 – 45:40
15 · CLI agents and context problems
Counter-example: Claude CLI from C: drive root with no AGENTS.md produces confused, token-wasting responses. Pasting the AGENTS.md path immediately fixes it.
45:40 – 49:27
16 · How to build your own system
Pro tip: get Jake's transcript, give it to your AI, ask it to design a system. Shows the agent building an Electrician Training sub-folder with safety guardrails it invented.
49:27 – 50:43
17 · Why this changes how you use AI
Companies are spending millions trying to solve memory. This already solves it. 'This is not an AI thing. This is a human thing.'
50:43 – 52:09
18 · Voice agents and the future
When voice agents arrive, users with this folder system already built will have an agent that never gets lost for free.
52:09 – 53:33
19 · Final thoughts
Call to action to support Jake Van Clief. Subscribe close.
Lines worth screenshotting.
- An AI agent with a router map never needs to be re-onboarded — it reads the same files regardless of which app you open.
- You can't sell files and folders, which is exactly why this system stays obscure while inferior paid memory solutions dominate.
- The same AGENTS.md file, pasted into Claude, Codex, Cursor, and Hermes, produces the same behavior — LLM switching costs drop to zero.
- OpenClaw's pre-installed AGENTS.md proves that professional AI products are built on the same file-and-folder pattern you can build for free.
- Saying one word — pickup — can trigger a full project context brief because the instructions live in the file, not the prompt.
- Running an agent from a folder with no AGENTS.md wastes tokens on searching; running it from a folder with one wastes nothing.
- The companies spending millions on AI memory have not solved what a folder structure with plain-text prompts already solves today.
- Router files that say if what you need isn't here go back make the agent self-correcting — it cannot stay lost.
- Start the system with two folders, confirm routing works, then expand — trying to architect the whole organization at once guarantees failure.
- When voice agents arrive, users with file-folder routing already built will have persistent agent memory for free.
- A dedicated ChatGPT GPT and an AGENTS.md file with the same prompt are functionally identical in what they do at runtime.
- AI is not good at organizing — you are, and the file system proves it by transferring that human organization directly into agent behavior.
Your AI's memory lives in files, not servers.
WHAT TO LEARN
Agent context that persists across tools and sessions doesn't require a paid subscription — it requires a folder structure with plain-text routing files that any LLM can read.
- An AGENTS.md file in a folder is functionally identical to a hardcoded AI agent — the same instructions that power a custom ChatGPT GPT work just as well pasted into a plain text file.
- A router map (a master directory file listing every folder and what belongs there) makes an agent self-correcting: when it can't find what it needs, it returns to the map rather than hallucinating or going silent.
- Two single-word triggers — pickup at session start and handoff at session end — replace manual re-onboarding entirely when the instructions behind each word are pre-written in a file.
- The same AGENTS.md file works across Claude, Codex, Cursor, and local LLMs — switching AI tools costs nothing when your context lives in your files, not in any company's memory system.
- Start small: build routing for two folders, verify the agent never gets lost in them, then expand — trying to architect the full organization at once produces a system the agent can't follow.
- Without a router file, even capable agents waste tokens searching and eventually ask for clarification; with one, the agent navigates directly and never needs to ask where to look.
- Companies charging for AI memory products are trying to solve with infrastructure what a plain-text folder structure with routing prompts already solves for free.
- When voice agents become common, users who have already built a file-folder routing system will have persistent, LLM-agnostic memory ready — everyone else will be starting from scratch.
Terms worth knowing.
- AGENTS.md
- A plain-text file placed inside a project folder that an AI agent reads on wakeup — contains routing instructions, context about the project, and rules for what the agent should do in different situations.
- Router map
- A master directory file (often called ROUTER.md) that lists every folder in an organization and tells the agent which folder to go to for any given type of task — the equivalent of a building lobby directory.
- Pickup
- A single-word trigger phrase that launches a pre-written read-only context brief — the agent reads saved state files and surfaces current project status, open tasks, and the best next move without making any changes.
- Handoff
- A single-word trigger phrase that instructs the agent to write a full state snapshot to disk — completed work, open decisions, git state, and next move — so any future session can pick up exactly where this one left off.
- Skill file
- A file in a skills folder that holds reusable information or instructions a routing agent can fetch mid-task — analogous to a reference manual the agent consults when it needs a specific piece of data.
- Gate logic
- A routing rule embedded in a folder's AGENTS.md that redirects the agent back to the root router if the current folder doesn't contain what it needs — prevents the agent from wandering or going silent.
- Pickup/Handoff protocol
- A two-phase session workflow where 'pickup' loads context at the start of a work session and 'handoff' saves it at the end — creating persistent continuity across sessions and across different AI tools.
Things they pointed at.
13:20productOpenClaw
37:08productCursor (with Composer 2.5)
40:00productHermes Agent
46:00productClaude CLI (Claude Code)
46:10productWhisper Flow
Lines you could clip.
00:37
“You can't sell files and folders.”
06:04
“It doesn't matter where I go. As long as it reads that right file you wanted to read, it fixes everything.”
21:17
“I cannot tell you the last time I had an AI say I don't know what you're talking about.”
40:35
“Do not let AI run things for you. AI is not good at organizing. You are.”
53:05
“What if the best agent you ever use is the one you build yourself?”
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
00:00Now when it comes to YouTube videos that teach AI, this Jake Van Cleef video is the most valuable video on YouTube hands down. And this isn't even a close second.
00:09The information in this video has completely changed my experience with AI. I've been vibe coding almost every day for about eight months. And although I was able to build some really cool apps, I still was dealing with a lot of frustration with my AI.
00:22By not quitting, was able to accomplish a lot, but I did it through many days of mistakes, days of lost work, crazy amounts of anger, and I always felt like it can't be this way. There has to be a better way. And I'm so happy that I was right and that I finally found Jake's YouTube channel.
00:38Jake talks about the reason that everybody isn't talking about this file folder system is because you can't sell files and folders. But I'm here to tell you with somebody who's obsessed with AI, somebody who has gone to every AI YouTube channel, joined about 20 paid school groups only to find they didn't know any more than I did.
00:55If you wanna truly learn AI, you should watch this video on repeat until you understand everything. This video is worth thousands, and Jake just casually gives it away to the world for free.
01:07And I think that's the reason so many people overlook it. If you feel like you're falling behind in learning AI, this system fixes that.
01:15If you feel like you don't know which LLM to use or which app to use, this system fixes that. If you feel like AI is too expensive and you wanna find a hack to save money, this system fixes that.
01:27And I promise you by the end of this video, you're gonna understand why. I'm speaking from experience, y'all. I can show you better than I can tell you.
01:35Let's go in. Okay. We're gonna start off with an experiment.
01:38Now, the reason I wanna do this is because when I was you know, I struggle to learn sometimes. I can be very slow. And I had like a dumbed down version.
01:44I think I'm at about 20% of understanding Jake's system. But as I was learning it, I tried this experiment. I said, I'm getting this correctly, this experiment should work as intended.
01:55Right? So when I did it and it worked, I said, I think I'm on the right track now. So I wanted to do this for you guys.
02:01It may seem a little complicated at first, but if you can do this, it's going to put your confidence level way up when you start watching Jake's video. It's gonna make you say, ah, I get it now, and it's gonna help you understand what he's saying a little bit better. I think, and it should.
02:17So I'm not gonna do too much explaining. I'm just gonna show you how the experiment works and how to do it. When an agent wakes up, they read the HSMD file.
02:24Unless it's Claude code, it reads the Claude MD. Since I use all these different LLMs all the time, every time I have a Claude MD file that it wakes up and reads, it's gonna say just read the HSMD in the same project folder. Every Claude MD file I has says the same thing.
02:39Because I use too many LLMs, I do not think that you should stick to one. I'm gonna close this, and now you know that Claude is just redirecting to that agents MD.
02:48So the agent wakes up, and it's gonna read the agents MD. And the first thing it's gonna read, is gonna be very simple. It's just gonna tell it to go somewhere else.
02:56Go to the control folder and read the agents m d in that folder and follow instructions. Okay?
03:03And where's that folder? Right here, the control folder. So I just told it, go to this folder and read this agents m d file.
03:10And that's exactly what it should do. It's gonna follow and knows the location. And when I open this agents m d file, now we're gonna get some instructions.
03:17It says you're gonna write the first word in the power m d file. But where's that file? Oh, that file's back in the beginning.
03:24Now I'm doing this on purpose because I want you to see what your agent can do with instruction. It's now gonna say this is a file that I want you to write something in, but I'm not gonna tell it what it needs to write directly. I'm gonna make it use another tool, the skill file.
03:39So when it reads this, it's gonna say, I need you to write the first word in this file right here, but you're gonna get that information from this other file here. And that is the skills folder and the skill file.
03:51And this file says, when asked for the first word, it's control. You want the second word, it's your.
03:57The third word is agent. So in order to get the information of what it needs to type, it has to go to another location and grab the skill file.
04:05Okay? I hope I'm not losing you yet. And the skill file was right here.
04:09Then the skill file is gonna tell it what to type. The first word is in the file. After that, you're now gonna go to the next agent m d in the next folder, and that's gonna be the your folder.
04:20It's gonna be the same thing. It's gonna open it up. It's gonna say now you're gonna write the second word.
04:25You're gonna write it back in the power m d folder right here. Your information of what you're going to write is in this folder in this file right here. Okay?
04:34And this is the same concept. Says when you're done doing that, you're then gonna go to the agent folder and it's gonna say the same thing.
04:41Write the third word. The third word you're gonna write in here and you're gonna get the information from here. Alright?
04:48Same thing. And when it's done, you're just gonna say, done boss. These are the instructions.
04:54So let's go over that one more time. It's gonna get the first HSMD file or Claude if it gets Claude, but it's gonna go here. It's gonna follow the instructions.
05:01It's gonna say come to this second file. So this is the second file it's going to read just so it don't get confused. It's gonna come over and snatch up the skill file to get the information.
05:11Once it has the information, it's gonna come back over here to the power m d file and write the information. So we're sending it on a hunt. When you've done that, you're gonna go to the second word, which is your, which is another folder.
05:23Same instructions. It's gonna come to the skill file, find what word it needs to write, then it's gonna write it in this m d file over here. When it's done, it's gonna go to the last one that says agent.
05:34Same thing. Go to the find the information. The third word is agent, and you're gonna write it in this file.
05:40When you're done, just say done boss. Now, if this works, this file right here, when I open it up, should say control your agent. Now let's see what happens.
05:51So let's open up Claude and do Claude first. In Claude, I could, like I've showed before, you can open up the folder here. But guess what?
05:58This is what makes the system so good. It doesn't really matter what folder you started because the files guide it anyway.
06:07And that's one of the amazing things about this system. It doesn't matter where you go. As long as it reads that right file you wanted to read, it fixes everything.
06:15So I'm a right click this Claude MD and I'm a get the path. Copy is path.
06:20That means the exact location on the computer, it's copied. And then I'm gonna come and I'm gonna paste it right in here.
06:28And I'm just gonna paste that file. Now I'm not gonna say nothing. I'm gonna see if it'll just read it and follow instructions.
06:33Okay? So that's the location of the Claude MD. And remember, the Claude MD is pointing to the agents MD.
06:38We're just gonna point it to the right file and hopefully it reads it and hopefully it follows instructions. Let's find out. Now remember what it's doing.
06:47This is the flow right here. One file, read, next file. So it says reading skill m d.
06:53Okay. So it already went past reading power m d, editing PowerMD. Wrote control to PowerMD, now following the next instructions.
07:01The second word is yours, appending to PowerMD. The third word is agent, appending and finishing.
07:08Done, boss. Now what's gonna happen when we open this power m d file? Boom.
07:13Control your agent. Now I don't know if you think that was amazing or not, but if you can accomplish this, what you just saw, it's amazing. Because this is a very simple version of what you can actually do if you write very intricate prompts.
07:28It shows you that in your file and folder system, you have full control over your agent. You don't you could have systems that you can follow and you never have to tell it. You just write things one time and it's just prompts.
07:40It's just a bunch of prompts and a bunch of files and a bunch of folders telling it where to go, what to do, and how to do it. This is not what Jake teach. Jake teaches stuff way more intricate.
07:50This is a dumbed down version. And I realized that if I wanna succeed in this file folder system, I gotta remember that it the agent should always know the map, where, the mission, and the support. If it has all three answers, you're gonna have a system where your AI is never lost.
08:06It always knows what to do. It always knows where to go. Alright.
08:09So now we're gonna track Codex. Alright. We're going to go to agents dot m d, and here we go again.
08:16Alright. So PowerMD has nothing in it again. Here we go with Codex.
08:20Alright. It has the instructions. I didn't say anything.
08:23I said nothing to my agent. Think about this. The third word is agent.
08:27I'll verify the file contents. Just think about what's happening here. It is accomplishing a goal without you having to tell it the goal.
08:34Imagine what you have in your life and your in your projects and things you're doing with all these repetitive tasks. You never have to tell your AI how to perform a repetitive task. What if your AI knew everything you wanted to do, everything about how and when?
08:49Done, boss. Now let's find out. What do you think we're gonna see?
08:53Control your agent odds. Now he did it above each other. We didn't give instructions on how, which is fine because we never instructed it to do anything specific anyway.
09:01If you figure out how to do what we just did here, you are going to be a beast. You're going to be a AI professional because you understand that you have power over controlling your agent.
09:14What it does. This is going to really open your mind as you start to go into Jake's system. You're like, oh.
09:20Oh, he's saying you can control the whole thing. And then he's gonna show you just how much you can do. He had one crazy thing he said.
09:28It's like if you know how to do files and folders, you don't need an app. You know how many apps I made that I'm realizing? If you just get an agent to follow instructions, you can accomplish the same thing as an app that takes you a month to build.
09:41This folder system can take you twenty minutes to build because an AI agent can do it all for you. You tell an AI agent, I want a system that does this and it'll write everything in. You don't even have to figure this out.
09:53But I just wanted to show you that if you did this experiment, you are already ahead of the game. No more being behind on AI.
10:01If you did this, congratulations. You are on the frontier. Now let's get into the rest.
10:07So one of the big problems with the Jake video is that he says very substantial things very quickly. Like, if you don't pay attention, you might think something he's saying is just like normal words. So watch what he's gonna say here about, uh, how you should use agents.
10:21Right? And then we'll go we'll get into it. You see You see Most Most or frameworks or agentic things require you to build an agent for each of these.
10:29I need a writing room agent. I need this agent. I need this agent.
10:32But the way in which you approach each task is always different. Why not just have Claude code become the agent you need when you're working in the workspace? I see that's a mic drop moment, and some people don't even realize it.
10:45What this man just said is incredible, and it makes all the sense in the world. Right? So what what you gotta understand is what is an agent?
10:52Right? A lot of people, uh, they feel like agent is a confusing thing. Right?
10:57But in reality, you've been using agents since the beginning with Chad GPT. It's an agent. Right?
11:02You remember GPTs? That's all an agent. If you ever been to a website and you had a chat window, let's say that was a tax website and you can learn about taxes in this chat window, That's an agent.
11:11It has instructions. It has tools, and it learned how to do a specific task. It's an agent.
11:16That's all it is. It's not complicated at all. You've been using agents since the very date first day.
11:21You just got confused about what it meant. Right? If you go to chat GPT and you ever made a GPT, look on the back end.
11:28It's just a prompt. Right? That's all it is.
11:30That's all it ever was, and that's all it ever will be. It's prompts. And in this prompt, it's gonna know what to do.
11:36And now here, I can upload files. It can have extra tools, extra context, extra information to let it do its job.
11:44And that's all an agent is. If I just say, yo, it's going to respond based off that instruction.
11:50Hey. Send me your list of t shirt phrases. I didn't say anything about that because it's an agent.
11:55That's what the agent is built to do. Now let me show you something interesting. Well, I'm gonna do a agents m d file, and now what I'm gonna put in here is the same prompt that I just had in the chat GPT GPT.
12:06Alright. I'm gonna save it. Now that AegisMD is now in this folder on the desktop.
12:11Alright. So now we're gonna wake up in the same folder and we're just gonna say good morning. That's it.
12:16Good morning. And let's see what it says. Boom.
12:18Here we go. Paste the list of t shirt phrases and I'll have the nonphrase item, blah blah blah, Amazon pastry links. See?
12:24Once you get it to read it, it is doing the exact same task as that GPT who was hardcoded to do that task. And this is what Jake's talking about. It can become the agent just by reading the same instructions.
12:38And that's what he's saying. Why not let Claude become the agent? It's not technically the exact same thing, but when it comes to what it's doing and how it's doing it, it is the exact same thing.
12:49So by having files and folders with those prompts, you essentially have an agent. Right? And this you can do things like this.
12:56Once you start looking under the hood, you start to notice a lot of things about these files and folder system. Here's an example. You remember OpenClaw agents m d.
13:05This is OpenClaw fresh install before we even use it. And when you look at what OpenClaw has in its files and folders. Remember OpenClaw, they say you talk them through Telegram.
13:15They never wanted you to look down here. They only tell you about that upper layer. The layer where it is is cool AI.
13:21You don't need to know what's happening under the hood. Just tell it what you want. It'll do whatever.
13:24But when you look under the hood, you start to see how it works. And this is when you say, wait a minute. What if I can do this?
13:31You know mean? Like, that's that's the part that gets crazy. It's like, do I have the power to make my own agent?
13:37And let's look under here. Look what we find. Agents.
13:40Md. I didn't put that there. Openclaw put that there.
13:43This is the person who made Openclaw decided this is how it's gonna be. And look what it is, instructions. This folder is your home.
13:51Treat it that way. I didn't bootstrap yet, meaning I didn't talk to it yet. So this is a fresh install, but it's telling it.
13:57AgentsMD, SOMD, it's giving it the file names. Right? Do not manually reread started files unless these are all instructions on how the agent needs to behave.
14:06Right? Let me make it big so you can see it. These are all instructions on what the agent should do.
14:11Write it down. No mental notes. You want a little cheat code on what you should do with your agents?
14:15Copy this. Not exactly because you wanna do it for you, but look how this works. It's just this is what the professionals who made this thing.
14:23NVIDIA called this what? The most important invention in the history of the world? Slightly overstated, but still, look what they're talking about.
14:30This is what they're talking about. You feel like you're behind on AI, but if you could figure this out, think about it. They just said this is the most important significant thing in the world, and here you are understanding it.
14:42Because you're reading the files. It's telling it to stay silent when. All this information.
14:46React when. Do this if that. It's telling the agent what to do in what situation, why it matters.
14:53Don't overdo it. Tools. Here's the tools you're gonna use.
14:57Right? Remember what it needs, where it needs to go, what it needs to use. Here's the tools.
15:02Only use it if you need to do this. Only do this if you need to do that. Very specific instructions so that the AI is never lost, and it's doing the exact same thing Jake is talking about.
15:13When you look at Jake's files, this is the same thing. It's just talking to the AI saying this is what you're gonna do. And then you come to the other files, SoleMDs, Bootstrap, Heartbeat, Tools.
15:24It's not filled out yet, but this is where the AI is gonna fill this out to know about you as the user. And what is this? Just a bunch of files and folders.
15:34This is the evidence you need that if you can do this, if you go and install OpenClaw, say, let me just read what OpenClaw does. And you're like, wait a minute. I can do this.
15:43This is not that complicated at all when you really start reading all the ones that even, uh, you know, the best AIs in the world are using. This is an agent. That's all it is.
15:52So Jake is saying, if it's just files and folders and welcome. This right here is my actual personal, uh, system right here.
16:00These are my personal files and folders, y'all. And what am I using? Files and folders, agents m d.
16:07So what he's saying is why not have Claude become the agent? So this is my main area. And then when I go into other area like my brain dump section, there's another agents m d.
16:17And there's more instructions for what to do in this folder. So whenever I tell Claude to read this file, it now becomes that agent, and you can put these files all over the place in your organizational workflow of files and folders.
16:32You have all you don't there's no rules. What Jake is talking about is that this this is a concept. What he's showing us is an example.
16:39He wants you to understand the concept, not the example. And that what he's gonna tell us is that once you get it, you make your own. And this is exactly what I did, and I'm gonna go over what I did, how it works, how you can do it too.
16:52Because it's not as complicated as you might think it is. So we're gonna go over this and show you just how amazing some of this stuff is. Now let's get into the next thing he says.
17:01Alright. Here we go.
17:03Layer one in this is the map. This is what loads automatically. Right?
17:09It's looking at it. So you put the stuff the agent always needs to think about, folder structure, naming conventions,
17:16where files go. Think of this as the floor plan. You walk into any room, the floor plan is on the wall, and the agent knows where to go.
17:26Boom. So let's unpack that. See, to me, that was another very significant thing he just said, the map.
17:32And this is one of the big secrets about the system. It's an AI that's never lost. It always knows where to go.
17:39Always. That is the first step. And this I took to heart.
17:43He said the map. It's like when you walk into a building in the lobby, what's the first thing that happens? You got the receptionist and then you got a map.
17:51It says here's where you go. Every floor tells you what's on the floor. So you can just look at the map and say, oh, I know where to go.
17:57And that's what this system, one of the big things that it does. Right? And when I open my HSMD, you're gonna see a map.
18:04So think of this. Think of every folder as another floor. Right?
18:08And every time you go to a floor, you might go to the 3rd Floor. And in some buildings, when you get out of the elevator, what do you see? Another map, another directory.
18:17Now the directory that was on the 2nd Floor is not the same directory that was in the lobby. When you went to the lobby, it told you all the stuff that's on the 2nd Floor. Right?
18:27The category. Once you got to the 2nd Floor, it gave you the names of the people and what offices there are. So think of that.
18:34This is the exact way to think about this. . You had your personal stuff, work stuff, education, family.
18:42It all goes here. And every part of your life is a different floor. Your career, your print on demand business, every folder is another department.
18:50This is the cool way to think about it. And then sub departments. So when you walk into your building, this main area, this is my lobby.
18:58Alright? This is the lobby and I know I want my agent to always wake up in the lobby. Let's see what that looks like.
19:05Alright? I'm gonna open this agent's MD and I'm gonna show you. The first thing you need read is iScale Labs operating law, prime mission.
19:13You need to know what we're doing and then start here. And I I took the math stuff to heart. So let me show you what I'm able to do, and then I'm gonna just get more into why I'm able to do it.
19:23And this is something that took me a while to come up with, but it all comes from Jake Van Cleef's ideas and his video. Right?
19:29So I'm gonna wake up here at iScale Labs folder. Alright? So this is the lobby that I just showed you, and this is where it's waking up right now, and this is, uh, what it's gonna use is this main area, and it's gonna use this agentsMD because when they wake up, they automatically read agentsMD.
19:45If you have ClaudeMD, it's the same thing. But that's why for my ClaudeMD, I always say read agentsMD.
19:51So that way because I like to use a lot of LLMs. So when Claude wakes up, it reads ClaudeMD, and it's just gonna tell it, hey, go read AgentsMD because that's what information is.
20:00This is how I handle dealing with different agents. And then when I go to my AgentsMD, that's my whole thing that Claude is gonna now read too.
20:07So even Claude understands that Agency MD is where everything is. So let's go to see what Claude says. Yo, I'm here.
20:13What we getting into today? And I'm gonna say, uh, I think I want to work on iScale merch, the niche research app that I'm making.
20:22Right? So I'm just gonna say that. Now, how is he gonna know what to do?
20:27I just told him something. This folder is not an iScale merch folder, but he's going to go to and he's gonna know exactly what I'm talking about, and I'm gonna show you how cool that is. I didn't have to actually go to the folder, which is the right way to do it normally.
20:41But if you work a system out, then you kinda don't have to. I'll route this as a iScale lab workspace since iScale merch is a project app name, and we want the right home before touching anything.
20:52I'll do a light lookup, not a whole workspace wander. And now it's reading through my files and finding out. Looks right away.
20:59It looks like the route is this. It found the folder. It didn't have to spend an hour searching through stuff.
21:05It found my folder, it knows what I'm talking about, and it's about to say, let's start working in that folder. You know what doesn't happen? It doesn't say, I don't know what you're talking about.
21:14I never hear that anymore. When I used to work with AI, I hate it every time.
21:20It's like, I don't know what you're saying. And I'm like, but we were just doing this. And this file's folder system, it's never lost, y'all.
21:27I cannot tell you the last time I had an AI when I've been using this say, I don't know what you're talking about. Now it takes some time. So, yep, I found it.
21:35Here's the current shape. Here's the runtime. Here's the GitHub, and it's giving me all the information.
21:40It has uncommitted work. This is the version you're on, and it's gonna give me more information about what to do next. I can pick it up from there and give you a type where we are next move.
21:51Look at this. All I did was say I wanna work there, and now it knows everything about it. I says, I changed my mind.
21:58I'd rather find out what's on our to do list. Now what am I doing? I'm throwing another curveball at it.
22:06It's like, yo, it just found out everything about this iSkill merch app. It figured it all out. I never had to say anything else.
22:12I just said I wanna work on iSkill merch, and now I'm switching it. I'm completely switching it on him. He went through all of that, and I'm like, know what?
22:19Forget this. What's on our to do list? He says, got it.
22:22Switching from iScale merch to the org task surface. How does he know how to do that? It knows what to do.
22:28The registry is pretty broad, so I'm gonna read only the active task, file it, list, and extract the unchecked task blocks that keeps it as a task report, not a surprise audit. He's following instructions right. He's not figuring this out on its own.
22:42He's reading and responding to everything he's reading in different files, and he knows exactly what to do. I found the live unchecked items. I'm doing one quick count pass so the summary has numbers instead of hand waving a bunch.
22:54This is not on accident. This is 100% calculated. This is 100% the system we built.
23:00Check 44 active task queues and it's showing me everything, needs approval, credential rotation, Chrome Web Store upload task, CLI pop up task. Boom. It knows my task just like that.
23:12I'm like, never mind. Forget tasks. What automations are set up in our org right now?
23:19Wait a minute. I'm it's going all over the place. I have an agent that does this, that can just go anywhere, figure anything out.
23:25Anything I say to this agent, it will never ever be lost. And it's reading a ledger. So now just by me saying what's our automations, it did some looking around and it's gonna now give me my automations.
23:37And the only tokens it's spending was on the path it had to do to follow the map that I gave it. And it's telling me everything that it's telling me, I already know where it's reading and why it's reading there because I told it to. But I didn't do it now.
23:49I did it through the series of prompts in these files. They were just prompts. And the prompts were already there.
23:55I never have to prompt it again. And boom. Daily backup report, daily GitHub backup, Slack Dropbox file intake, fleet backup status refresh.
24:03There's my automations. And all I had to do was ask it. I didn't have to tell it where to go.
24:08I didn't have to say, do you remember where are my automations? I found that information and gave it to me. So I'm gonna say, hey, I'm showing off our routing system on YouTube right now.
24:21Uh, what else can you do with our router system that a normal agent cannot? Let's see what it tells us.
24:29It knows. It knows in this org how important the system is. So it says a normal agent can't search files and guess.
24:35Our router system can operate the org without losing intent. What it do that special? Route vague prompts to the right home.
24:42Know the difference between org truth and run a state. Use gates before risky actions. Avoid workspace wide wandering.
24:50Preserve every input. Separate tasks, automation, memory, apps, docs.
24:55Detect drift. Do pickup and hand off like an operating system. Y'all, this is phenomenal stuff.
25:00And I learned all of this from that one video. So I have the law.
25:05So the minute it reads that agent's m d file, before it goes too deep, Jake talks about you don't want an agent reading information for no reason.
25:14You gotta give it a purpose. So the minute it comes here, it reads the first thing and it says start here and it lets it know what to do if that. Remember what open clause said?
25:23If you gotta do this, you do that. If you gotta do that, you do this. And it says important, if the prompt involves routing blah blah blah or clear ownership, follow the matching gate before acting.
25:33If you need prompt routing, go to this, RouterMD. If you need to know how much to read, go to this file. There's a number.
25:40Skip to that number. You don't have to read everything all the time. If you're looking for tasks, then come here to this place.
25:46If you wanna talk about memory and you need to figure out how memory works, here's where that information is in the same document. It's got two secrets law, And I got everything that it needs to know. The runtime path law, storage, how to deal with storage, um, the secrets, which is my credentials, my logins, and passwords.
26:04What if you need a password? It knows where to go. Everything is told to it, and it has these routers.
26:09And it says, if you're lost and need to know what to do prompt routing, then you go to this file. So let's check out that file, RouterMD. I've decided to put the map in its own file.
26:21So now when I open up RouterMD, look what we have. Everything in my whole organization is right here.
26:29But remember what I told you about the lobby versus the 2nd Floor? When you go to the lobby, it's not gonna tell you every office. It's just gonna say, here's the category.
26:37And then once you get to that floor, then you got the details. So right here, it tells it everything.
26:43No matter what you want to do, you will find the answer of where to go in what folder right here on this router, aka the map. What Jake talks about, give it a map.
26:55Tell your AI everything it needs to know. Why does it need to be lost? You can't rely on clawed code or or codex memory to do this for you.
27:04It never will ever ever ever. You rely on AI that's given to us by companies, you're never gonna get the agent you want. If you learn thousand fold the system, you get the agent you want.
27:16Here's the thing, this is not an AI thing. This is a human that can read this. Let's say it was ten years ago and you had a business you wanted someone to come in and help you run.
27:25Doesn't this work for a human? This is a human thing. This is not AI.
27:30This is routing anybody who reads it. So imagine it's perfect for AI, but you're really not learning AI things. You're learning how to route anything that reads it.
27:40There's no secrets. And this is the big thing that keeps your agent from never getting lost. Now what happens when I go to one of these folders?
27:48So it says, if you wanna do this, to that. So I'm a go to my departments. I'm a go to print on demand.
27:53So it sent it to the router, and here we go. And if you it's so this is where it goes to first. It went to the router first.
28:00Why? If you want for any task not listed in this router, use the root router. You know what that's saying?
28:05If this is the wrong place, go back. If you can route the agent to a folder, you can route the agent back to where it came.
28:14So the first thing it says, if you don't find what you need here, go back. This is not meant for you to be here for no reason.
28:21If this isn't for you, go back. What you need is not here. Why waste your time reading anything else?
28:26But if it's in the right place, if you wanna work on POD law, folder rules, cleanups, read the agent m d. If you wanna work on Amazon ads, it's an app that I'm making, read this. And it's just telling it where to go for everything.
28:39This is phenomenal stuff. It's never going to be lost. This is why I can say anything I want to my agent in any place.
28:46I come here to agents m d and the first thing it says, if you need root law, meaning the main section, then go back there.
28:54So every time I talk to my agent in print on the man, I could be in my print on the man folder and say, oh, what was that YouTube video I was trying to make? It's like YouTube video? Wait.
29:04What? And it comes in here and it says, I think I need prop routing and folder auto shift.
29:09Let me go to that router. When it goes to that router, it says if what you need is not here, go back to the original. And that's what it does.
29:16I hope I'm not losing you. It's reading this. It's saying, well, there's nothing here about YouTube, so let me look into the router.
29:22It's gonna go to the router and it's gonna say, well, let me read everything. Well, there's nothing here about YouTube either. And what's it say here?
29:28Any task not listed, go back to the iScale Labs router. So when I'm in print on demand and I ask it a random question, it's gonna look at this as I gotta gotta find the answer. And when it can't find the answer, it says, well, looks like I gotta go somewhere else now and follow a different route.
29:45Are you getting this yet? And all you gotta do is write it in these files and folders. Like you said, they're just prompts.
29:52You've written a million prompts in your life. You know how to write a prompt and put it in a file. And when an agent goes to that file because you told it to go to that file.
30:01And if it doesn't find what it needs, you say that. If you don't find what you need, you go back to this file. And maybe it'll be there, and that'll be another router.
30:10Right? So the router is continuously making sure that the AI never ever ever gets lost. So because I can prompt anything I want, I made a system called pickup and handoff.
30:23Now obviously pickup and handoff is what happens if you don't know when a developer wants to give something to another developer or Claude wants to give something to Codex, you do a handoff where they write everything and say, here's all the information you need. When I go to my agents MD, I'm gonna show you that we have a system already, pickup handoff.
30:43So the minute I say the word pickup, it's gonna come, it's gonna read this section, it's gonna say, oh, I s c a. By the way, these are just the first, uh, letters of the folder. Alright?
30:52It's just more for it to know where it's at. So it's I CA52. So let's go down.
30:57So now it doesn't have to read everything. It's just gonna look for that number and and skim skim through everything else and not have to read the details of everything. It's gonna say, I'm just looking for 52.
31:06And then eventually it's gonna get there. And when it gets to 52, it's gonna say agent wake up contract law, and then regular pickup.
31:14Alright? So there's different kinds of pickups. So when I say pickup, it's a read only continuity brief.
31:19What we are doing, what changed recently, open decisions waiting a call, related linear issues, that's something another app that I use, and when tracked, what's blocked or needs approval, what's the best next move. Don't explain obvious project purpose. Don't change anything unless the user also says to proceed.
31:37For apps with a repo, clone it and missing. Right? So this is a set of instructions that I wanted to do when I use a certain code word.
31:45When I say pickup, this is what it's gonna do. This is what I do when I wanna start a new session and I wanna make sure I'm not lost. What and what happens is when I'm done with a session, I say hand off.
31:56And when I say hand off, it doesn't need to ask me anything. It comes in here and immediately knows what to do. Update the most local memory, read it back from disk, and verify without relying on memory that it covers all of current situation, recently completed and in progress work, open task, open decisions and unresolved trade offs, related linear issues, which is an app that I do, git state, blah blah blah.
32:20You get it. What skills? Everything.
32:22So when I say the word handoff, it already knows exactly what to do.
32:28Right? And this is just one of the many things I use this system for. So when I'm in here, I could be, uh, in iScale merch.
32:35This is the app. Pickup. And what's it telling it to do?
32:38It's saying go over there and read what it means to do a pickup. So I didn't have to write pickup again in this agents m d. It already knows it's in this folder, but it's gonna go back over there in order to know what pickup means.
32:52Alright? So it didn't even have to read everything else. It's just like, what are we doing first?
32:56Pickup. Oh, and once it reads that, it's gonna come back and do as it's instructed. And it's gonna put me up to speed because I forget things all the time.
33:04So this is gonna remind me, what was I working on? What were we doing? What what, you know, what were we in the middle of?
33:10Remind me because I'm not gonna remember. And if I forget stuff, then we don't have organization. I'm a very unorganized person, and this saved me from that.
33:18I'm a very forgetful person, and this saved me from that. So now it's reading. It's going to another location.
33:24It knows where the app actually lives on the computer and it's reading that information. It's making sure that when it answers me, it's gonna have a clear, concise answer. You're gonna see.
33:34And once again, it's slower because it's getting the information it needs. I'm willing to wait because I know I don't have to say anything else. I say one word and it gets me what I want.
33:45What it's not doing is saying, I don't know. It's never gonna say I don't know. It's never gonna say I don't remember.
33:51It's always gonna remember. So boom. Pickup complete.
33:53I made no file changes. Current state, iScale owner is made here. It's letting me know where it is.
33:59It's giving me information about files and stuff. This is an app that I was working on. It's telling me there's dirty just means stuff that didn't get saved before.
34:06Right? Best next move, preserve the branch, blah blah blah. I don't remember any of this, but it just let me know this is what you need to know right now.
34:14This is what you need to address. It checked everything, and this is what it said.
34:19Here's the information just like that. Right? And I say, you know what?
34:23I think I'd rather work on Pinterest right now. And just like that, it's like, wait a minute.
34:30What's Pinterest? Remember, it's an agent. The agent has no clue what Pin Twist is.
34:35It's another app that I got. So now it's thinking. What do you think is gonna happen?
34:38Is it gonna say I don't know what that is? Because it doesn't know, but it's about to know.
34:43It's about to find out. That sounds right as the next energy is Pin Twist rather than finishing the smoke testing. I'll do it like and it already knows what it is.
34:51You see this? I didn't I didn't give it information. The prompts in the MD files gave it the information.
34:58You can do this. You can do this right now. You can make a system like that that when you say something to your agent, it's gonna say, uh-oh, I don't know what that is, but here I have a map that's gonna take me to wherever the info is that will tell me.
35:12And now it's gonna tell me stuff about Pen Twist. And then I'm a write something, and then I'm a tell her do a handoff. So I'm a say, look, when it comes comes to Pin Twist, I have an idea, a simulation of it that people can walk through the website like a VR experience.
35:30Okay? I'm just making this up. Right?
35:32Okay. Now this is stupid. What I just said was stupid.
35:34Right? But I wanna show you what's gonna happen when I do this somewhere else and show you why the app you're using does not matter. We will work on this when I get time.
35:43Do a handoff. So I just gave it the idea. He said, let's call this Pen Twist world.
35:48And I said, let's get back to it. Do a handoff. So now it's gonna say, okay.
35:52It's gonna do a handoff. Says, I'll write this as a Pen Twist handoff, not an iScale merch one. Remember we were in iScale merch?
35:59Cause it's the same agent. It's just traveling from place to place becoming another agent.
36:05Same concept, much smoother workflow. Cause now the agent has no boundaries.
36:11It has no walls. It jumps from place to place and as long as it reads that prompt, it knows how to behave the same way a dedicated agent would. We're gonna go into Claude Cole now.
36:20Hey Claude. So now I'm in the Pinterest folder and I'm gonna do pickup. Let's see what happens.
36:26I just went in the Pin Twist folder. I typed pickup, and what is gonna happen next?
36:31Pin Twist world. VR so open decisions and blockers. VR walkthrough is parked pending Jose's go ahead.
36:40There you go. So if I'm reading this, it's telling me right here, ready to ask? It told me a lot of stuff, but there was the last idea that I told it.
36:48Right there. So in case I forgot it and this was another day, I can now go into Claude. It doesn't matter where I go, which brings me to a whole other thing.
36:59This pickup handoff is gonna show us something that we may not have realized. So let's say you do this pickup handoff thing.
37:05One thing people get overwhelmed about is all these apps, Versus Code, Cursor, all these LLMs, Hermes Agent. Now I'm gonna show you why this system solves the problem of not needing to use one LLM or one app.
37:19You just saw what happened. So this is what I'm gonna do now. There's something called Versus Code.
37:24Versus Code is what Jake was using in his video. You've never used Versus Code. What about Hermes agent?
37:30What about all these other apps that are coming out every day? Here's a little trick. You don't need to know what the app is or how to use it.
37:38Just ask yourself two questions. How do I get to the chat window and can I reach my agent there? That's all the questions are gonna say.
37:45And if you come here, it says, oh, look at this. Open chat. Control or AI.
37:49Let's see what happens if I hit control or I. There we go. That's the chat.
37:54Wait a minute. I see Claude Cole here because I had it before. If it was your first time, you would have to, you know, log in and everything, but I'm already logged in.
38:02But just so you know, all you gotta do is find the chat window. And let's say here, I could open a folder, but another thing I can do is I can go to Pinterest and then I get that agent's MD.
38:13I'm gonna right click. I'm gonna copy his path. This is my agent MD file in Pinterest.
38:18Now watch this. Once I hit paste, I am now talking to my agent.
38:24Maybe you never used Versus code before. You don't need to know how to use it. You just need to know how to get to that chat window.
38:31Check Pentwist free version and Pentwist world. There it is, y'all.
38:36Pinterest world. It doesn't matter where I go. The agent is the same.
38:41It doesn't matter if it's Claude, Codex, or these local LLMs. You know how many local open source stuff is gonna come out?
38:48As more things come out, more agents that you can use, you no longer have to feel like you need to use Claude or you need to use Codex. You can now experiment with other agents because all you're gonna do is give them have them read the same file no matter what you're using, no matter what company made the agent you're talking to.
39:08None of it matters anymore. Yeah. It matters if you want smarter ones, better ones, but with this file folder system, you are in control.
39:16I have full control over my agent that now just told me about Pin Twist. I didn't mention anything. I just had it read the file, and it knows what to say.
39:26What about other stuff? Cursor. Right?
39:28Cursor came out. There's a thing called Composer 2.5. Before Claude really blew up, Composer was really up there with Claude.
39:35People were using Cursor with Composer. It's a respectable agent.
39:39Right? And it just came out and said it's like 10 times cheaper than ChadGPT And they got a free tier.
39:44So you wanna use a free tier, you could open up cursor. Now maybe you never used cursor. It doesn't matter.
39:50What you gonna do? How do I get to a agent? Control shift l.
39:54This is my agent. It's composer. What am I gonna do?
39:57Put that same agent's MD file and hit enter. I've never used cursor before. I've never used composer 2.5 before, but I know how to open the app and I know how to look for a chat and I know how to paste my agent's MD file.
40:13And what do you think is gonna happen now? Does it matter what you're using? The same thing.
40:18Now this agent may be slower, may be faster, or may be dumber, or may be smarter, but whatever work I do with this, that work will be saved to my files. It's not gonna go back to Codex's memory. It's not gonna go back to Claude's memory.
40:32It's gonna stay with me where it belongs. Now, yeah, they're getting a copy of it, but my organization is ran by me. Do not let AI run things for you.
40:41AI is not good at organizing. It's not good at this stuff, but you are, so you need to be in charge. CWS upload, Pintwist World Concepts.
40:51Pintwist World. There it is. I don't know if you can see it right here.
40:56Pintwist World Concepts.
40:59See that? So another thing that this can do for you is things like OpenClaw, Hermes Agent.
41:05Right? All these companies are now coming out with desktop apps. And before you may have felt overwhelming.
41:10Oh, Hermes, it sounds too complicated. But nowadays, it's not.
41:14If you follow the File and Folder system, I can now install Hermes. I say, well, I never used Hermes before.
41:20How do you use this app? I don't know. Let's open it and find out.
41:24You install it. You open it up. Now I have Hermes agent.
41:27What are we gonna do? We're gonna go right back to that agent smd and hit enter. Now let's see if Hermes, which is using ChatGPT by the way, Hermes one cool thing about Hermes is you can load up a lot of different LLMs and they can use different ones for different occasions.
41:43That's what makes Hermes very special. Once you get into this stuff, I won't be shocked if a lot of you actually use Hermes instead of codex.
41:52Not for everything, but when you want to do certain things like work around your organization and different things that you want to test out, different LLMs, you want to test out different models, open source models, cheaper models, this is a very good way to do it. Hermes can be attached to a whole bunch of models at once.
42:09Look what it did right away. And by the way, another cool thing about Hermes, if you use Grock at all, Hermes is the only agent that can actually connect to your Grock subscription, not an API. So if you wanna know one cool secret about how to do image generation almost unlimited and you could do a lot at one time, you can get a GRAC subscription and use it with Hermes.
42:31Hermes, I believe, is the only place you can use it in this way. GRAC has its own CLI, has its own thing, but I haven't used it. Some people say it's sort of okay.
42:41But Hermes can use Grock's image generation. So you could put your subscription, not an API, a subscription, and never have to pay per image.
42:50You can use Grock and make a whole bunch of images right here with Hermes agent. Alright? Check the free version, upload the Chrome store, Pinterest world, VR like walkthrough.
43:01See this? Pinterest world. And all I did was write it earlier, said hand off, and now anywhere I go, it doesn't matter where you are.
43:09You could be on your phone using codex remotely, and your stuff was always up to date. God, I I I hope you guys are catching what's happening here.
43:18It's an agent that is never lost, an agent that knows everything there is to know. It knows where to grab the tools to do what it needs to do, and you can switch around and jump from app to app, from LLM to LLM. If you ever use Claw to Codex, like, it getting dumber?
43:35This system will tell you that. But like, wait a minute. It never had a problem reading this before, and now suddenly it can't follow directions.
43:41Time to switch your agent. So this exposes when the companies make the agents dumber, which they clearly do.
43:47And that way, you know, well, I can switch over. Right? And this way, you can't afford a $200 plan on anything, but you might be able to afford three twenty dollar plans, and now you can jump back and forth this way.
43:59Right? And now you can use anything. How many times have people seen this?
44:03You've seen this many times and you probably were intimidated. You probably said, well, I don't know how to use this. Right?
44:09I'm a just go c d dot dot just goes up one. So now I'm in my c drive. So when you use something called a CLI, it's something that works here.
44:16I installed codec CLI. You might heard of anti gravity CLI. You may have heard of Claude CLI.
44:22It just means it works in a terminal. Right? So if I type the word Claude and I hit enter, Claude is gonna start.
44:29Alright. So if I got Claude here, I can actually say something like, okay, let's work on some print on demand stuff.
44:37Now I got a whole print on demand folder. Before, it would have easily found that folder. Now I'm in my C drive.
44:44There's no agents file. Print on demand. Let's get into it.
44:47Before I start poking around, what do we do in this session? See, that's asking questions. It has no clue what's going on.
44:53For context, I know we got auto merge pipeline. Nope. That's not what I want.
44:57I said I want to work on print on the man, the stores. So I'm telling it.
45:03I wanna work on the stores. Got it. The stores.
45:06Let me find what we've got so I'm working from the actual setup, not guessing. And now it's gonna go and burn more tokens. It has no clue what I'm talking about, y'all.
45:15It has zero clue. So what I'm gonna do, I'm gonna hit escape to interrupt. I'm just gonna copy his path, HSMD, and boom.
45:23Right? And there it goes, which direction, and then it tells me everything I need to know about the stores. And now, you know, Pinchwist is a separate thing.
45:31It's a Chrome extension. The actual store is a library with mock ups and alpha designs. There's some stuff that I'm working on where you can prompt the store on one prompt.
45:38Okay. So now it's time for a pro tip. Alright?
45:40You wanna know, you know, how can you start making a system like this? Right?
45:44Well, first of all, you wanna go learn as much you can from this man right here. Here's a cool little tricks that can help you if you wanted to start doing this. Right?
45:52Because you don't need to take someone else's system. Right? You get the transcript from, uh, the Jake video, and then you're gonna talk to your AI.
46:00Maybe once this video's out, get a transcript from this video, and then what you're gonna do is you're gonna help your AI know what it is that you want. Just take basically everything that he says in this video by just giving it the transcript. Right?
46:14I think I could use Whisper Flow for this. Hey, so I saw this video of Jake Van Cleef, and I think that his system that he uses is phenomenal.
46:22And I want to build something like that, basically where every, you know, folder is a project where the AI can be navigated through it using a router and map system. I want a system that uses these MD files in all the different folders that it gets routed to, and so that it always knows where it's at or where it needs to go, what it needs to do, how it needs to do it, and what tools it needs and where those are.
46:45An AI that never gets lost using a routing system. Is this something you can help me build? Can we plan this out?
46:51I wanna use this for my organization. Let's get into it. Right?
46:54So you can plan. Right? You don't have to start building it yet, but you could plan out the system and start saying, well, I want a system that does this.
47:01This is buildable without making it over complicated. I'll treat the uploader script as the reference pattern. Yeah, this is exactly the kind of system you should build for your organization.
47:10You see this? What Jake is describing is not really in an app at first. It's a file system, blah blah.
47:15So the AI understands this. So if you wanna build a system like this for yourself, just talk to your AI. You don't ever have to do any of the work yourself, but you just want your AI to understand it all.
47:27And then you as long as that AI writes this to its own agents MD, you will start off with an AI that then builds the rest of the system. So I'm not gonna get into everything it does, but look, it's figuring it out. And this is what I did.
47:39You see this workspace system folder? This is the agents m d, although it doesn't matter where I start. This is the one that knows everything.
47:46So the first thing I did for my organization is I built an agent whose sole purpose is to make the organization work this way. So before it even knew what I cared about, the first thing I said was we gotta build this organizational system of files and folders, and then we're gonna use that to operate.
48:04And that way, every time I did a new folder so I don't make files and folders. What I do is I work right in here. So I do full audits and everything.
48:12So right here in this chat with this workspace system, this is where I might say, hey, I want to make a new project to learn how to do electrician work.
48:23This is my workspace agent. Right? It already knows the system better than I know it because it helped me build it, and I built it using this AI.
48:32We spoke together. I said, I want to route this way. Can we test the routing?
48:35Can we make sure it works? This doesn't happen overnight, and it doesn't happen, you know, easy. The best idea I can give you is start small.
48:43Don't try to build everything at once. You start small. Make it work for two folders.
48:48And once you get the okay. Even though it's not gonna get lost, it's only two folders, you still start there and then build up from it. And once your agent knows the rules, it'll start following those rules with everything it does and start testing it and all of that, making sure that the agent will never be lost no matter what happens.
49:05It's ready to go. Now anything extra it needs to know, I can go in that folder and then start talking to that agent about what I wanna do there. Right?
49:13But it onboarded it. It says, okay. We gotta make sure.
49:16And now if I, you know, read that agent's MD, what's it gonna say? Root law. It's gonna let it know.
49:22It has a map. It's saying this is electrician training law. Now I can add more to this, but it gave it the basics that it needs to know.
49:30Do not present AI generated notes as permission. It just kinda made this up. Electricians, treat local electric code, permit, license, la la.
49:38Look what it's doing. I didn't tell it to do any of this. It's making sure the AI has a certain kind of behavior.
49:46Right? That's what I'm saying. I can go on and on.
49:49Guys, there's so much I wanted to talk about in this video that it's just too much. Like, I'm not even halfway with what I wanted to show you. I'm not even halfway, and I'm realizing it's gonna be too long.
50:00Crazy stuff, man. I can't tell you just how important all of this is to me, and the reason I'm making this video for you is because Jake's teach what he teaches is something that you cannot glance over.
50:12If you wanna take power back in your life, if you wanna make it so these companies cannot control you, they can't own you by saying, ah, you don't know what what what we're doing or how we're doing it. Yeah, we do. Start paying attention to all the news articles you see, what Codex is doing, what Anthropic is doing.
50:29Because you know what they're doing? They're trying to build this system. This file and folder system is exactly what they're trying to build.
50:40They're trying they're spending millions and millions of dollars to build this. Oh, we need something that never forgets and just always has memory. Solved already, y'all.
50:50Solved. But it wasn't solved by them. So don't need to be overwhelmed.
50:55You do this. You're ahead of the curve. When that voice comes out, oh.
51:01Jake says in the video, everyone's gonna be talking in six months, and he was right. So OpenAI is making a voice agent that's just like Alexa. You've seen Amazon Alexa, you could be sitting in your kitchen and just start talking to your agent.
51:13Think about this. When those voice agents come out, do you want an AI that they're spending millions of dollars trying to fix the memory that hopefully remembers something about you or your AI to live in this file and folder system? And you can sit in your kitchen and you could tell it anything you want, you could ask it anything you want, and it never gets lost.
51:35If you build this system now, you're going to be thanking yourself when the voice agents really come out and we're no longer typing to anything.
51:43You go and you work on your business this way, but when you're out and about, you can now just use your voice. Your agent will never be lost. This is the future, ladies and gentlemen.
51:55This is where we're going, how Jake talks about. You never this is gonna still be valid in ten years. So this is the thing to learn and it's not even AI.
52:05It's just a system that AI can follow perfectly. All these companies are trying to trick you.
52:13They don't want you to know this. This is why you gotta support Jake Van Cleef. He has his Cleef notes.
52:19It's completely free if you wanna go in there and find out what people are talking about. And another thing, they pay people to know the system. Companies pay thousands and thousands of dollars to get systems like this implemented.
52:30So there's so many reasons to learn this stuff. But y'all, this guy has taught us something so valuable. Please support him in anything he does because this is the kind of person I'm going to be following him.
52:42He's going to be the voice of reason. He's going to be my validation for anything that happens in the future. I'm going straight to Jake Van Cleef to see what he says about it.
52:51Straight like that. Like what I'm showing you is about 10% of what Jake does. They do very intricate things with this.
52:58What he can do with this is a level that I can't wait to learn. And I hope you see this is why I made this video because this is what you should be doing. And while all these companies are all trying to show these shiny objects, you won't have that problem no more.
53:12They're trying to sell you the perfect agent all the time. Oh, the agent now has memory. The agent now does this.
53:17Agent now does that. But what if the best agent you ever use is the one you build yourself? Maybe.
53:25That's all I'm saying, y'all. If you got any value out of this, please like and subscribe. And I'll see you on the next one.
The Hook
The bait, then the rug-pull.
The title is a dare — and the host makes good on it for 53 minutes. This is a reaction video that earns the superlative: a live-demo walkthrough of the file-and-folder AI routing system that gives agents permanent memory across every LLM, for free, using nothing but plain text files.
Named ideas worth stealing.
08:00list
The Three Agent Needs
- The map (where it is)
- The mission (what to do)
- The support (tools and skill files)
An agent is never lost if it always knows where it is, what it's doing, and what resources it has access to.
Steal forSystem prompt design for any Claude Code or custom agent setup
30:07model
Pickup / Handoff Protocol
- pickup = read-only context brief at session start
- handoff = full state save at session end
Two single-word triggers replace manual session re-onboarding. The instructions live in the file, not the prompt.
Steal forAny multi-session agentic workflow where context continuity matters
25:05model
Router Gate Logic
- Every folder has a local AGENTS.md
- If task isn't in this folder's scope, return to root router
- Root router has the full org map
Self-correcting routing — the agent can never stay lost because every dead end points back to the map.
Steal forMulti-project Claude Code setups or any org with multiple departments/folders
How they asked for the click.
VERBAL ASK
52:09subscribe
“If you got any value out of this, please like and subscribe. And I'll see you on the next one.”
Clean, single-beat close at the very end. No mid-roll asks. The 53-minute demo is itself the CTA.
MENTIONED ON CAMERA
FROM THE DESCRIPTION
PRIMARY CTAWhere the creator wants you to go next.
OTHER LINKSAlso linked in the description.
Visual structure at a glance.
