With memory switched off in both apps, the same prompt run three times against the same folder shows Claude following the rules and ChatGPT ignoring them, until it gets pushed.
Posted
yesterday
Duration
Format
Demo
educational
Views
2.2K
35 likes
57 · 43
Big Idea
The argument in one line.
Memory settings don't decide whether an AI assistant reflects your standards; a structured local folder of contracts and named sub-agents does, and in this test Claude follows that folder's rules more consistently than ChatGPT.
Who This Is For
Read if. Skip if.
READ IF YOU ARE…
You run recurring AI workflows (content, LinkedIn posts, client work) and want the same quality every time instead of re-explaining your standards in every chat.
You already maintain, or want to build, a structured local folder of instructions and specialized sub-agents that multiple AI tools can read.
You're deciding between relying on a model's built-in auto-memory or building your own external memory system.
SKIP IF…
You only use AI for one-off, single-prompt tasks where auto-memory or no memory at all is already good enough.
You want a beginner walkthrough of prompting — this assumes you already run Claude Code or a CLI coding agent day to day.
TL;DR
The full version, fast.
With memory and custom instructions switched off in both apps, the video runs the same LinkedIn-post prompt three times: with no folder, with a shared folder open as a ChatGPT project, and again with that folder in Claude Code and Codex CLI in the terminal. Every time, Claude reads the folder's AGENTS.md contract, takes the orchestrator role, and routes the task to a named specialist sub-agent that follows the creator's actual teaching; ChatGPT produces a generic post that ignores the folder's content unless pushed. The conclusion: auto-memory is fine for simple asks, but a folder of contracts, an agent index, and layered guidelines/SOPs/workstreams is what makes AI output repeatable across models, and now that AGENTS.md is a shared standard, the same folder works with Claude, Codex, and Gemini, so switching models the moment one underperforms costs nothing.
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.
Tom opens both desktop apps side by side and turns off every memory feature in both: Codex instructions, Codex memory, and Claude's chat search/reference and memory generation, deleting stored local memories too.
01:16 – 02:42
02 · Test 1: same prompt, no folder
With no folder or context, he sends the same LinkedIn-post prompt to GPT-6 Astra (medium) and Fable 5.1 (medium). Both answer fast; Fable's is longer and adds hashtags ChatGPT didn't.
02:42 – 05:36
03 · Test 2: my folder as the context
He repeats the prompt, this time pointing each app at his 'My Life Folder - TR' project, the free ICOR-for-Life scaffold containing AGENTS.md, CLAUDE.md, and an AI Team of named agents.
05:36 – 07:11
04 · Claude reads agents.md and delegates
Claude announces itself as Larry the orchestrator, reads AGENTS.md, and routes the task to Sage, the LinkedIn specialist sub-agent, visibly launching and completing the sub-agent run.
07:11 – 09:46
05 · The two LinkedIn posts side by side
Both drafts share the same three hashtags and no emojis, but only Claude's post reflects Tom's actual ICOR teaching about calendars versus task lists; Claude also raises a 'DECISION NEEDED' block about scheduling.
09:46 – 10:56
06 · Test 3: Codex vs Claude Code in the terminal
To rule out the desktop app's harness as the variable, he reruns everything in the terminal: OpenAI Codex (gpt-6-astra medium) on one pane, Claude Code (Fable 5.1, high effort) on the other, both pointed at the same folder.
10:56 – 13:15
07 · Orchestrator and sub-agents in the terminal
Claude Code again identifies as Larry, states the scope, and hands off to Sage as a background sub-agent he can click into and watch work; Codex's result stays generic while Claude's matches Tom's teaching, though slower.
13:15 – 14:11
08 · Why Claude sticks to the rules
When pushed, Claude Code names its own miss out loud, admitting it gave Sage a generic brief without checking it against Tom's teaching, then produces a corrected draft grounded in the ICOR source material.
14:11 – 14:53
09 · Auto memory vs a local folder
Tom's thesis: auto-memory is fine for simple, repetitive asks, but breaks down once work spans multiple structured workflows, which is where a hand-built local folder of contracts and named agents pays off.
14:53 – 16:20
10 · Inside the folder: agents.md and the orchestrator
He tours the folder root: AGENTS.md, the root orchestration contract that used to be CLAUDE.md and is now also picked up directly by Claude the same way Codex and Gemini read it.
16:20 – 18:02
11 · Agent index, guidelines, SOPs, workstreams
Inside 06 AI Team he opens agent-index.md, which tracks agent hires and retirements, Charta's AGENT.md (infographic designer), the numbered Guidelines library (GL-001 through GL-031), and hundreds of numbered workstream files.
18:02 – 19:22
12 · Why this saves tokens
Because guidelines, SOPs, and agent files are indexed and only pulled in when a task needs them, the orchestrator loads a small slice of context per request instead of one giant instruction file.
19:22 – 20:33
13 · Why I stick to Claude
Tom says he's run this folder-based system over a year across model updates, and Claude has been the most reliable at proactively following the folder's rules without repeated reminders.
20:33 – 21:32
14 · Where ChatGPT still wins
He concedes GPT-6 Astra produced better, faster visual and slide-deck output in a separate test, while the Claude-based agent overthought the task and took longer, hence 'most of the time.'
21:32 – 23:24
15 · One folder, any LLM: the free scaffold and the course
Because Claude now reads AGENTS.md like Codex and Gemini, the same folder works across all three. He points to the free downloadable scaffold and the paid myICOR course/membership, then asks viewers about their own memory approach.
Atomic Insights
Lines worth screenshotting.
Switching off every memory and custom-instruction setting in ChatGPT and Claude isolates a fair test: whatever an AI 'knows' about you has to come from a folder, not from prior chats.
Given the identical prompt with no folder open, ChatGPT and Claude both produce fast, generic LinkedIn posts, showing the base models alone don't carry a creator's specific teaching.
Pointing both AI apps at the same project folder doesn't guarantee the same output: only Claude reliably reads AGENTS.md and follows its orchestration rules by default in this test.
Claude Code and the Claude desktop app both self-identify as an orchestrator named 'Larry' who is contractually barred from doing task work himself and must delegate to a sub-agent.
ChatGPT's output on the same folder stays generic ('you can use the calendar if you have a small workload') while Claude's output repeats the creator's own rule almost verbatim.
Running the same test a third time in a bare terminal, Codex versus Claude Code, rules out the desktop chat UI as the reason for the difference; the gap holds in the CLI too.
When challenged, Claude Code names its own mistake out loud ('I gave Sage a generic writing brief without requiring a check against your teaching') rather than repeating the generic draft.
The folder splits context into three tiers: guidelines (numbered standards), SOPs (step-by-step procedures), and workstreams (numbered execution logs), so only the slice a task needs gets loaded.
An agent-index.md file tracks a full hire-and-retirement history for the sub-agent roster, treating removed agents as retired rather than deleted.
A monolithic single-agent setup that holds every instruction in one file tends to work well at first, then degrade as the workload gets more complex, which is the problem this layered folder is built to avoid.
In a separate slide-deck test, GPT-6 Astra produced better visuals faster, while the Claude-based agent overthought the same task and took longer, which is the reason for 'most of the time' rather than 'always.'
Because Claude now reads AGENTS.md the same way Codex and Gemini do, one folder of contracts works across all three, so switching models when one underperforms requires no rewrite.
Takeaway
A folder of rules outlasts any one model.
SYSTEM OVER MODEL
Reliable AI output at scale comes from a structured folder of contracts, named agents, and layered references that any model can read, not from picking the single 'best' model.
01Memory off in ChatGPT and in Claude
Before testing whether an AI 'knows' your standards, turn off every memory and custom-instruction setting in the tools you're comparing, including chat search/reference and auto-generated memory, and delete any stored memories too.
A fair before/after test needs a genuinely blank state on both sides; leaving even one memory toggle on lets a model quietly carry over context that isn't coming from what you're actually testing.
02Test 1: same prompt, no folder
With no folder or context provided, both a leading Anthropic model and a leading OpenAI model produce a fast, generic answer to an open-ended prompt; speed and superficial polish don't indicate the model understands your specific standards.
Differences you see with zero context, like extra hashtags or more detail, are cosmetic model habits, not evidence one model is better at following your rules.
03Test 2: my folder as the context
Pointing an AI app at a project folder is only useful if the app actually reads what's inside it; opening the same folder in two different tools doesn't guarantee the same behavior.
A folder built for this purpose should hold named specialist agents, not just general notes, so a request can be routed to the agent whose job that specific task is.
04Claude reads agents.md and delegates
When an orchestrator role is defined in a contract file, look for the AI to name that role, state the scope of the request, and explicitly say who it's delegating to, before doing any of the actual work itself.
A visible handoff, where a sub-agent is launched and then reports finished, is a signal the system is following its own design rather than improvising.
05The two LinkedIn posts side by side
Surface-level similarity, like matching hashtags, length, or no emojis, can mask a real gap in whether content reflects your specific teaching versus a generic version of the same topic.
A well-designed folder should also produce structured decision points, such as 'approve this draft for scheduling?', instead of just handing back a finished asset with no next step.
06Test 3: Codex vs Claude Code in the terminal
Repeat a comparison in a second environment, like desktop app versus terminal, before concluding a difference is about the model rather than the interface wrapped around it.
Opening a CLI session directly inside the target folder, rather than the root drive, is what makes sure its available context matches the folder you intend to test.
07Orchestrator and sub-agents in the terminal
A CLI agent that lets you click into a running sub-agent and watch its own transcript is more auditable than one that only shows you the final answer.
The same generic-versus-specific gap that showed up in the desktop apps reappearing in the terminal is the actual evidence that the difference is the model's behavior, not the app.
08Why Claude sticks to the rules
When output doesn't match a folder's actual teaching, push back explicitly and ask whether the instructions were read; a well-built system should be able to name its own miss rather than defend the generic version.
An orchestrator admitting a specific, checkable mistake is a stronger reliability signal than an answer that merely sounds more confident.
09Auto memory vs a local folder
Built-in auto-memory is adequate for simple, repetitive requests where a small amount of remembered preference is enough.
Once work spans multiple structured workflows with their own standards, auto-memory's 'AI decides what to keep' approach stops being reliable, and an externally defined folder of contracts becomes necessary.
10Inside the folder: agents.md and the orchestrator
Keep the root instruction file focused on the orchestrator's role and boundaries; use it to point to the rest of the system rather than trying to hold every rule directly.
A shared root contract file that multiple AI tools read the same way is what makes a folder-based system portable across models instead of tool-specific.
11Agent index, guidelines, SOPs, workstreams
A single index file tracking every agent's hire and retirement date turns 'who does what' into something the orchestrator can look up instead of something you have to remember.
Splitting reference material into standing rules, step-by-step procedures, and execution logs keeps any one file from becoming unmanageably large.
Numbering reference files makes them easy for both a human and an AI agent to index and retrieve on demand.
12Why this saves tokens
Loading every instruction into one giant file costs tokens on every single request, even the ones that don't need most of that content.
An orchestrator that only pulls in the specific guideline or procedure a task needs keeps context small and keeps quality from degrading as the system grows more complex.
13Why I stick to Claude
Judge an AI tool's reliability on repeated, structured work over many months, not on a single demo; a system holding up across multiple model updates is stronger evidence than one good answer.
A model can appear not to follow instructions in isolated complaints while still being comparatively more rule-following than alternatives when tested side by side on the same task.
14Where ChatGPT still wins
Don't treat one favored tool as universally best; a competing model can still outperform it on specific work, like faster and better visual or slide-deck generation.
The honest framing is 'most of the time,' not 'always'; naming the exception you've actually observed makes the overall claim more credible.
15One folder, any LLM: the free scaffold and the course
Standardizing on a shared root contract format means a single folder-based system can serve as insurance against any one model regressing or falling behind.
The goal of a system like this is to be able to switch the underlying model the moment it underperforms, without rebuilding your instructions from scratch.
Glossary
Terms worth knowing.
AGENTS.md
A root instruction file, read by Codex, Gemini, and now Claude, that defines an AI's role and boundaries and points to the rest of a project's context.
Orchestrator
The single AI agent a user talks to directly; it is barred from doing task work itself and instead reads a request, decides scope, and delegates to specialist sub-agents.
Sub-agent
A specialized AI instance launched by the orchestrator to do one job, such as drafting a LinkedIn post, governed by its own AGENT.md contract file.
SOP (Standard Operating Procedure)
A step-by-step document describing exactly how a recurring task should be carried out, stored separately from the top-level instruction file and loaded only when needed.
Guideline (GL-0xx file)
A numbered reference document defining a standing rule, such as file naming or a design system, that agents look up instead of having it duplicated in every contract.
Workstream (WS-0xx file)
A numbered log file recording how a specific piece of work was carried out, used as institutional memory the agents can consult later.
myPKA
The creator's term for the personal knowledge-and-agent system this folder structure implements, taught in more depth in his linked course.
ICOR methodology
The creator's productivity framework for planning time and tasks, used in this video as the test for whether an AI actually reflects the creator's own teaching.
“The calendar holds constraints. Meetings, appointments, dates other people set. Things you cannot move.”
crisp, definition-style line that teaches the core ICOR distinction in one sentence→ IG reel cold open↗ Tweet quote
11:48
“The list stays the single place where work lives... The calendar is for time, the list is for work.”
quotable punchline that resolves the whole test in one line→ TikTok hook↗ Tweet quote
12:17
“I gave Sage a generic writing brief without requiring a check against your teaching. That was my mistake.”
a rare on-camera moment of an AI agent naming its own specific error→ newsletter pull-quote↗ Tweet quote
16:06
“Larry, the orchestrator, is not allowed to do any work at all.”
counterintuitive rule that explains the entire system design in one sentence→ TikTok hook↗ Tweet quote
The Script
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.
17px
In this video I want to talk about personalized memory of your AI agents and the LLM that you are using. And here you see on the right side it's the desktop application of Claude and on the left we have the desktop application of ChatGPT. And here I am in the settings and the codex instructions are empty.
This means there are no instructions that it will pick up the moment I start a chat. And also the codex memory is switched off and I even clicked here on delete to ensure it's really empty. And here on Claude you see the same thing i switched off both right there's no search and reference chats there's no generate memory from chats because i don't want ai to randomly generate memory from different contexts so here ai is deciding what to keep and what to ditch and so on instead we are building this memory locally ourselves and i'm just showing you here that it really doesn't matter if you're using chat gpt or cloud especially with the latest update in cloud where now picks up agents .md instead of claude .md.
What this means, we will look into this, so don't worry if you never heard about this. In this video, you will learn everything about memory and why you might also want to keep this switched off.
So now that we have both memories switched off, let's go to a new chat and let's create something. This is the prompt that we will use. create a linkedin post about the pros and cons of managing your tasks directly on a calendar we will use astra medium to really show you that it doesn't matter how good the model is we will run into the same issues okay on the left we have chat gpt selected to work and on the right we have claude in the co -work section here so as you have seen in my previous video chat and co -work merge together into one thing and in order to make it equal to the work section in chat gpt we will switch to cloud code here and do the same thing in cloud code on the desktop application.
And here we ensure that there's no folder selected. So there's no context, nothing. Here we will pick Fable and we will use medium effort.
And then let's see what it will come up with by using just a prompt without any knowledge. Okay, now we see the outcome and you see Fable went into more detail than JetGPT. We also see here that it added hashtags while we don't have any hashtags here but that's not the point i'm going to for here that could be due to the model or it could be just by accident that here we have this versus here we're having this outcome but it was very quick in both sides Now, let's do the same thing.
Let's do a new chat here, also a new session. This time, we choose a project as they call it here, which essentially is picking a folder. And here I'm using my own my ICO for life scaffold that I have in here with my personal information.
And as I'm using this to build our membership platform, and this is all about productivity, and we have the ICO methodology where we describe how to manage your time and how to plan out your tasks, I'm expecting now that AI will go in here to find better context proactively, I will use exactly the same prompt, but proactively finds more information to make this LinkedIn post much more sound like myself.
So here, it's important to see you see here these agents .md file and the cloud .md file. And if you're following us, you know that we are providing this free scaffold where this is actually leading to a team where a lot of team agents are in here in my case they are doing all different jobs and there's even a team member called sage and he's the linkedin specialist and i'm expecting now that the ai will work according to this scaffold and if you look into these instructions this is just triggering making the ai working as larry an orchestrator so he's just picking up understanding what i'm asking for and then forwarding the work to specific agents who are doing this job then because they have their own agent .md file to give them instructions how they should work.
So here we will already see by just opening up this folder in here, and here I still need to create projects to load this in. Here I can just simply pick the folder and it will look into it. We will see what's happening.
And then I will do one more test that will probably give us even deeper insights into how everything works. The point is we talked about memory in the beginning. I switched it off to make clear that nothing is picking up from previous memory how I worked with AI.
And it's all on a blank plate. All it has now is the contracts inside this folder to understand how to work on something in there. And it's all based on these agents.
.md files. And the news from Claude is now that Claude picks up the agents .md file directly and no longer needs to rely only on a Claude .md file. So this means agents .md file is what usually ChatGPT and Codex was looking up in any folder to see what this folder is about and what our instructions are out of the box.
That's an OpenAI thing. And Claude is now adapting to also use agents .md files. So that makes our folder that we use use here, even more LLM agnostic.
And I can really point both LLMs on the same folder and they pick up the instructions. However, let's see if... both LLMs actually act the same way.
And we see already the difference here. Here we have already the outcome. I'm routing this to Sage for clear balance draft.
So it picked up the agents .md file and it recognized that he's the orchestrator and he will look into who is available. And then he launched this sub -agent, as you can see, where the sub -agent read this agent .md file. I can open this up and here are the instructions what the sub -agents should do as a LinkedIn expert.
And then he created this. post based on these instructions then he read it he sent a message to sage again and then he again there's a back and forth so that's clear orchestration going on here and here's the outcome and it looks already completely different than what we had before with the emojis and things like this fable also finished now and you can see here also the agent that ran and this is larry the orchestrator who launched this sub -agent and sent this prompt to the sub -agent to work on And here again, read your contract at agent .md.
See, that's exactly what I just showed you in the folder. And then he provides what I want. But you see already, I didn't change anything from the prompt.
I said, create a LinkedIn post about the pros and cons of managing your tasks directly on a calendar. No mention of ICO or nothing. But here you see Thomas, the co -founder of the Peoples Movement, my ICO, a productivity learning platform, and so on.
So he picked up already the context, the orchestrator picked it up and shared the essential context that the sub -agent to create a LinkedIn post actually needs in order to make this agent highly efficient in focusing on creating this post. And this is why we have now these two posts available.
And you see, both don't contain any emojis. Both contain now hashtags and even the same three hashtags. And you see the length is equal.
Well, and the approach and what is written inside these posts differs. But I have to say, when I read through this, this is much more generic and leaning towards, yeah, you can use the calendar and if you have a small workload and so on, while here it goes much deeper into what our actual thoughts about these are, because here's the general rule that we say, the calendar holds constraints, meetings, appointments, dates, other people set, things you cannot move.
task manager holds freedom everything where you decide when so task list is infinite and a calendar day is finite and that's why we plan out the day and time block bigger blocks on the calendar where we do our deep work shallow work i'm not going too deep into this because that's what we teach in our icon methodology inside the courses but here it's much more reflecting what we actually teach versus here it's kept much more generic and here it gave me even more additional information and And here, this is a decision block that's based on instructions that I have inside this folder again, that whenever he wants me to make a decision about something, he should provide it this way so I can just use this code and answer here if it should get posted.
And here you see this decision is that he looked up about what's the general work that we do on a daily basis. And he clearly understands here that we have Buzz who is connected to our social media scheduling platform. Should he actually schedule post should he create this it's a follow -up quest and i said not for today just a test and here it just gave me a result and just one thing one outcome so like maybe it is the thing that you're asking for and i have to say many times fable and no matter claude in general is too wordy it's just listing me too many things making too many follow -up questions where it asks me things that i constantly think okay i could do this too and i constantly get into to this run of, yeah, let's do another thing.
And yeah, he's probably right. We should do this too. Versus here, I asked a question and I got the outcome.
And you see, these are the top models, GPT -6 Astra and Fable 5 .1. They're both the top models from both these LLMs and yet both are pointed to the same folder and yet we get completely different outcomes.
So let's do one more test because what we are running in here is a harness in a desktop application that might change things from one LLM to the other. So that's why I want to switch to the terminal and do the exact same thing again. You can simply do this by going to any folder that you think there's context in there that you want to work with.
And you can right click and you can say new terminal at folder. And this means that this opens up in your terminal. See there's on top the folder.
whatever I do in the terminal will apply to this folder. Another way would be I can create a new window. See, this is now on root.
There's no folder, but I can say CD and drag and drop this in and hit enter. And now I'm also in this folder. Okay, there are many ways.
So both terminals are now in these folders. On the left one, I will launch codecs, which is the chat GPT, the CLI, see OpenAI codecs. And I have here GPT -6 Astra medium.
And here I will launch Claude. And here we are.
This is now Claude. And this is on Fable 5 .1. Now let's see if we get any difference.
Here's the same prompt. And here the prompt again. And let's go.
As you can see, he immediately says, I'm Larry. Scope. What is the scope?
And he finds Sage. And what I love, he launches the subagents down here. I can click the subagent here.
And I see now he sent the prompt again, read your agent .md file. And then this subagent is right running. now.
And I can visually see, okay, I'm in a sub agent and here's my main session. So I can perfectly switch between the sub agent. That's really something I really like.
And he already comes up with the result here. And again, he comes up with the real advantages if I put it directly on my calendar and it's, it's still the more generic thing. So now I could force it to say, did you actually read the instructions inside the folder that you're working in the agents .md file and so on, because this output doesn't seem to really reflect what we while here it takes a lot longer than what it took.
Here, and yet, the outcome you will see is a lot more detailed. And it feels like that he really took the context from my folder where my whole life is in there. And it's all the things that he needs, including my tone of voice and all these things.
And here we go. Important thing, what I do now. The list stays the single place where work lives.
Only a few things get a calendar slot. Deep work, thinking, anything that dies when the day gets loud. Everything else sits in the list and gets pulled into the gaps.
The calendar is... for time, the list is for work. So that's perfectly what we say.
This time, he picked up other hashtags. Let's see what hashtags we have here. Well, here, there's not even one hashtag, even that Sage, the sub agent for LinkedIn is instructed to act.
But he says, I gave Sage a generic writing brief without requiring a check against your teaching. That was my mistake. Okay.
And here we go. The missed distinction is, and he missed it twice because you have seen in a desktop application, it also didn't stick to the rules and didn't go too deep. So I feel it's using a lot more tokens, obviously, but it's much more sticking to the rules.
If I would go to Opus, I would get similar outcome as we have here. As you can see, there's also the decision making that is coming up that's based on my custom instructions that I have. And it seems that there are no decisions to make.
So no question about Buzz posting later. No decision make. So the thing is, I'm using over a year all these different LLMs and doing much more complicated stuff than just writing a LinkedIn post.
And that's where I'm trying to show you here one example of many endless use cases where I saw that Claude is just sticking to the rules much more. Even that some people complain, my Claude is not sticking to the rules that I gave it and so on. And yet compared to other LLMs, it's a complete different behavior.
And there it's not really the model doing it. Yes, if I would use Haiku, it might also go off rail. And I made a video about this already where I showed you the comparison between the different models and the different output that we get from it.
But as you can see, when it comes to the general proactive usage of rules, Claude was working out of the box correctly versus JetGPT didn't. And here we see now he uses the same symbols and everything. And now he provides it this way.
And now we get the same question. So my point of this video is actually to making clear that if you have the auto memory switched on, it will just generally keep in mind what you're telling. If you keep repeating things and you're annoyed about certain things that might work for simple work, but if you actually get work done and you're working on many different things that have specific workflows that you need to work through, like content creation, for example, this is the moment where a local folder in the right structure really shines and then you have these specific agents working for you that are all just based inside a folder the only goal is that we can guide the ai from one point to the other and that's how it's happening Yeah, it goes into the folder, the scaffold that you can download.
It looks into the agents .md, which was previously the cloud .md, but now is the agents .md for both, for chat GPT and for cloud. And here you see what this is all about, that this is a root orchestration contract and what you should do and what he is, the harness layer, many things that are stored in here. And that's not something I have written manually.
This is this auto memory, right? Whenever I say keep this in mind, he's not keeping this in mind. in a random cloud folder, it is keeping it in mind inside my own folder the way that I want it.
I say update your agents .md file in order to keep this in mind. And the thing is here, this is the starting the main session AI agent. that we call Larry in here.
He's the orchestrator. He's the single point of contact. He's the only person I'm talking to, and I expect him to delegate the work to other agents.
Now, in this example, you have seen that he's been off a sub -agent Sage. However, I have endless example where I work, there are six to seven different sub -agents all working in parallel, while still being able to talk to Larry at the same time. And in my opinion, just by what I showed you here, that you can switch between the sub -agents, subagents.
You could even write to these subagents. This is much more transparent in the cloud version compared to the chat GPT version. So the contract here says that Larry, the orchestrator, is not allowed to do any work at all.
If I want to create a slide deck, for example, he would go into this AI team folder and then he goes into the AI team subfolder and here's the agent index. And in here, there are all the agents listed that are available and what they are capable of. And there, Therefore, he would look through this index and he will find charter.
She is. this one we have even profile images for all these and this is her agent file and in here it's explaining you an infographic designer and what she's doing and all the instructions how we expect these things to do and that's where everything is interconnected because not everything is sitting in this agent .md file how things work because in here you see then things like this but those values are pulled from gl003 which is guideline 003 and that's where this agent then looks up the guidelines and these guidelines live in the AI team knowledge.
So this means in here we have guidelines, we have SOPs and we have work streams. And if you look into this, you see in my case, there's a lot of different work streams and these work streams point into SOPs, standard operating procedures, where there are step -in -step explanation, how things are carried out and so on. And you see there are a lot of things that I created over the years with all the different agents and the different use cases.
that I'm using it for. And then we have here the guidelines. And if you look into GL003 design system, this is the design system that describes how we design our membership and our branding.
The incline branding is described in here in detail. And again, it's indexed. It's easy for the AI agent to find what he needs to look up, but look at the size of these things.
So if I would load all these instructions all into one agent .md file, would be an endless document and a lot of context that the AI would need to load in immediately when you start the session. Instead, we start the session and Larry, the orchestrator, picks up based on this what I'm asking for and then looks into the details to find out what I actually need and only loads in the guidelines, the SOPs that are needed or the sub -agent actually does that is needed to get this specific job done.
And this is actually token. saving instead of having everything in one agent that gets more and more complicated you might have recognized this using other agents where you have one single agent and you give him all the work and it works very great in the beginning but the more complex and different things you do it starts more and more to fall apart or no longer stick to the rules and that's why this agentic work here comes into play where it's important that all these different agents here's for example bus as you can see here's the avatar here's the agent and all these agents also have their own journals where they create journal entries based on feedback that i gave so they can look up when they work with something specific and they run into issues how did we solve this in the past and this is all running on autopilot that's why it's so important to me that i'm using the llm that works most as an orchestrator with multi -agents projects and doing project work and claude is going more and more into the direction maybe we end up that
Claude out of the box inside the harness will take over of this. However, we have this already for over six months working this way that they now switch to agents .md if something as you can see that we always did because the Claude .md file was just pointing to the agents .md file. And if codex comes in here, he reads the agents .md file.
And if Gemini comes in here, it reads the Gemini, which also points to the agents .md file. And these agents point then to the rest of this folder what this is all about. So as you can see, everything has its place and it's only picked up once it's needed.
And I think this is an important understanding that many people don't have. They watch videos and they say, oh, Astra GPT -6 and so on is so amazing. And again, kind of one prompt games and so on.
Yeah, that might be the case if you do one prompt things. But if you do daily work on a recurring basis and you want to have the same quality level outputs, this is what you need to keep in mind. And this is what you need then to compare.
And that's what I did. I used Astra GPT -6 to create a slide deck that Faber had to feed. just didn't get complete.
And I realized now that Astra makes much better visual representations or animations than Claude did so far, even that they are very specialized agents and so on. First of all, it took much longer for Fable to create something because it's too specific and it's overthinking very quickly. And Astra, as you have seen, just throws it at you and you say, okay, that's it.
And then you have to follow up rather instead of getting too much back. So we need to find a sweet spot and this is where we can go even wilder where i could say now to claude as the orchestrator to launch a sub -agent that is actually using gpt codex gemini and whatnot to generate something and that's what i actually have when i generate images obviously claude doesn't have an image generator but claude can now launch sub -agent who uses then an api connection to generate images for example and this is how once you understand this basic functionality that we teach much greater in a course and this is where you can by the way download the folder for free and if you want to really understand how all this works we have a full course where we go into detail how this folder is structured how everything is interconnected and if you go through the course you see here step by step how we build up these folders and the different files and what is contained in the files there's a video for each step and there's a magic slide explaining exactly how this agenda team is working and once you
have this your llm agnostic you are independent because you understand how the agentic system is built which is the my pka concept that we are teaching inside my icore and once you understand this let's say claude falls short in the future then you can quickly and that's what i do i can switch to another llm i can use whatever i like at the same folder and see does it actually work better or not and that's something also whenever new model updates came out and people complained it's not behaving completely different at least these guidelines and guardrails kept my team moving forward properly so yes there were fluctuations and a bit differences between the models but in the end i still got the same results maybe it took longer sometimes it was shorter sometimes i needed to remind it and it was writing it down into the instructions and so on but this is the way i do it and i do it for over a year this way and no matter what model came out it remained the same system as you can learn
inside this course and inside our membership. Let me know in the comments below, are you using this system too? Do you use a local folder memory or do you fully rely on the auto memory updates that you can get in the different LLMs?
Or maybe you have a complete different approach. If you find this interesting and you want to see more videos like this, subscribe to the channel so I can catch you up in the next.
The Hook
The bait, then the rug-pull.
Tom switches memory off in both ChatGPT and Claude, then runs the identical prompt through three separate tests, with no context, with a shared folder open as a project, and again in a bare terminal, to see whether either model actually reads and follows what's inside that folder.
Frameworks
Named ideas worth stealing.
05:36concept
AGENTS.md root orchestration contract
A single root file that assigns the AI to an orchestrator role (named 'Larry' here), forbids it from doing task work itself, and tells it how to find the rest of the folder's context.
Steal forany recurring AI workflow where the same quality needs to repeat across many sessions and tools
16:44list
Agent index + per-agent AGENT.md files
agent-index.md roster with hire/retire history
one AGENT.md contract per named sub-agent
avatar image per agent
A roster of named specialist sub-agents (Sage, Buzz, Charta, and others), each with its own contract file describing identity, role, and operating principles.
Steal forsplitting one overloaded AI assistant into task-specific specialists
16:40model
Guidelines / SOPs / Workstreams layering
Guidelines (GL-0xx) — standing rules
SOPs — step-by-step procedures
Workstreams (WS-0xx) — execution logs
A three-tier knowledge structure that keeps most context out of the top-level instruction file until an agent's specific task actually needs it.
Steal forany large personal or team knowledge base an AI agent has to navigate
The ICOR teaching used as the test: a calendar holds fixed commitments other people set, a task list holds everything where you decide when, and only a few things should ever get a calendar slot.
Steal forany time-management or productivity content, and as a check for whether an AI actually retrieved a creator's specific teaching versus generic advice
CTA Breakdown
How they asked for the click.
VERBAL ASK
21:36link
“By the way, you can download the folder for free, and if you want to really understand how all this works, we have a full course.”
Delivered as a natural extension of the demo itself, this is literally the folder just shown working, rather than a separate sales pitch, with the free download positioned ahead of the paid course and membership.
Add Modern Creator as a preferred source and Google shows you more of our breakdowns in Search, Top Stories, and AI Overviews. It only changes what you see, and you can undo it in your Google settings anytime.
Add to Preferred SourcesOpens your Google source preferences with us pre-loaded. Tick the box and you're done.
A 34-minute tutorial that open-sources Tom Solid's myPKA scaffold — one markdown folder that any agent CLI initializes itself into and turns into a 4-agent PKM team.
A 35-minute live demo of a local browser app built with Claude that replaces Notion, Sunsama, Heptabase, and every health tracking subscription — all from one markdown folder.
00:26
00:43
01:01
01:36
01:54
02:11
02:29
02:46
03:21
03:39
03:56
04:14
04:32
04:49
05:07
05:24
05:42
06:17
06:34
06:52
07:27
07:45
08:02
08:20
08:37
08:55
09:12
09:30
09:47
10:05
10:23
10:58
11:33
11:50
12:08
12:25
12:43
13:00
13:36
14:11
14:28
14:46
15:03
15:38
15:56
16:14
16:49
17:06
17:24
17:41
18:16
18:34
18:51
19:09
19:44
20:02
20:20
20:56
21:12
21:47
22:05
22:24
22:40
22:58
23:15
34:33
35:38
09:09
09:02
21:38
13:10