Big Idea

The argument in one line.

Most AI tutoring fails because it is stateless — every session forgets you — and /teach fixes this by writing your entire learning arc to disk so the agent always knows what you mastered and what to teach you next.

Who This Is For

Read if. Skip if.

READ IF YOU ARE…

A developer or technical creator using Claude Code who wants to learn new frameworks or tools with real structure instead of one-off Q&A.
Someone who has tried asking ChatGPT or Claude to explain a concept repeatedly and noticed the answers drift, repeat, or forget context across sessions.
Anyone working through a technical topic — open source models, a new language, LLM internals — who wants a tutor that tracks actual progress rather than starting from scratch each time.

SKIP IF…

You need hands-on physical practice that cannot be captured in code exercises or HTML checklists.
You are already expert-level on the topic; the skill is calibrated for genuine learners with a real gap, not for quick expert reference.

TL;DR

The full version, fast.

The /teach Claude Code skill by Matt Pocock turns your agent into a stateful personal tutor: it writes a teaching workspace to your file system — mission statement, numbered lessons, learning records, glossary, and resources — so the next session always starts from where you left off, not from scratch. The underlying philosophy is a Knowledge-to-Skills-to-Wisdom ladder: lessons pull from vetted sources, require you to actually do something, and eventually hand you off to a community so you stop depending on the AI. Every lesson is calibrated to the zone of proximal development — just hard enough to grow, not so hard you quit. The whole thing installs in one line and starts a new learning track in about five minutes.

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 →

Chapters

Where the time goes.

00:00 – 00:50

01 · Intro

Teases the lesson output on screen, names the skill and its creator Matt Pocock, previews the three sections of the video.

00:50 – 02:00

02 · Stateful vs Stateless

The core concept: every standard AI answer is stateless like a substitute teacher; /teach is stateful like a real tutor who remembers you. Illustrated with an animated slide.

02:00 – 03:20

03 · Teaching Workspace Structure

Tours the actual file system: lesson plans, numbered learning records, mission.md, glossary, resources.md. Shows the folder in VS Code.

03:20 – 06:50

04 · What A Lesson Looks Like

Walks through lesson 1 (run first model on own machine with Ollama) and lesson 2 (hardware VRAM calculator for model selection). Shows interactive HTML lesson format with checklists and self-checks.

06:50 – 07:50

05 · Pocock's Philosophy

The Knowledge-to-Skills-to-Wisdom ladder: vetted sources, required practice, community handoff. The slide diagram shows the progression.

07:50 – 08:30

06 · The Learning Zone

Zone of proximal development: each lesson is calibrated to be just hard enough to grow without overwhelming. Illustrated with a three-tier diagram (Too Easy / Just Right / Too Hard).

08:30 – 09:05

07 · Setting You Free

The skill hands you off to a community once you have mastered enough — explicitly designed to break AI dependence, not extend it.

09:05 – 09:30

08 · Installing /teach

One-line install: paste the GitHub link into Claude Code, agent installs the skill. No manual steps.

09:30 – 12:15

09 · Full Demo

Live demo: runs /teach on 'how to build LLMs,' answers onboarding questions (why + depth), watches workspace build, discovers a path error, fixes it. First lesson generated: LLM as next-token predictor.

12:15 – 12:45

10 · You Still Do The Work

Reframes the tool honestly: it is a tutor, not a cheat code. You still have to do the reps. Understanding cannot be outsourced.

12:45 – 13:05

11 · Free Resources

CTA for free School community (AI Automation Nexus) with resource guide, install commands, and example prompts.

13:05 – 13:35

12 · Outro

Subscribe and comment CTA. Daily video pledge through July 11.

13:35 – 15:10

13 · Blessing

Kacper's signature closing: personal prayer for the audience. Optional — he invites viewers to click off if uncomfortable.

Atomic Insights

Lines worth screenshotting.

A stateless AI answer is a substitute teacher who forgets you the moment the session ends; a stateful one remembers exactly where you got stuck.
Most AI tutoring fails not because the AI is bad at explaining things, but because every new chat starts at zero.
Writing your learning arc to the file system — lessons, records, glossary — is what makes an AI agent genuinely teachable to work with, not just informative.
The zone of proximal development is not a fuzzy educational concept; it is a measurable calibration that makes the difference between a lesson you do and one you abandon.
Every /teach lesson ends with a required action, not optional reading — because action is the only thing that moves knowledge into skill.
The skill's goal is to make you good enough to join a real community and stop needing the AI — it is explicitly designed to make itself unnecessary.
You can outsource research to AI, but you cannot outsource understanding: the reps still have to happen in your head.
A vetted source embedded in a lesson plan is worth ten random answers from a chat session — provenance matters for retention.
The /teach workspace self-documents: your own questions and stuck points become the raw material for the next lesson's calibration.
Five lessons in, the agent already knows your hardware constraints, what you have tried, and what vocabulary you are missing — no re-explaining required.
Interactive HTML lesson files with live calculators and quiz forms beat static text explanations because they force engagement instead of passive reading.
An AI that eventually pushes you toward a community of real practitioners is more trustworthy than one that tries to keep you dependent on it.

Takeaway

What stateful learning actually changes about how you grow.

WHAT TO LEARN

The reason most AI-assisted learning does not stick is not the AI — it is the lack of memory: every session resets, so every session re-explains.

A tool that writes your learning arc to disk — lessons, questions, stuck points, glossary — compounds across sessions the way a real teacher does; a chat window resets everything.
Requiring action at the end of every lesson is not a design choice, it is the mechanism: knowledge without a task attached stays inert and does not transfer to skill.
Calibrating difficulty to where a learner currently is — not too easy, not overwhelming — is the single variable that determines whether a lesson gets completed or abandoned.
Sourcing from vetted, high-trust references and citing them in lesson files is what separates structured learning from the hallucination-prone answers of an open chat session.
The goal of a good tutor — human or AI — is to make itself unnecessary: when the learner is ready to join a practitioner community, the tutor steps out of the way.
You can outsource research and explanation to an AI, but understanding only forms through the reps you do yourself — the AI can structure the practice, not replace it.
A path error during a live demo fixed with one plain-English instruction illustrates that the agent is orchestratable, not fragile, when you treat it like a collaborator rather than a search engine.

Glossary

Terms worth knowing.

Stateful: A system that remembers context across sessions — recording what happened previously so it can act on that history in the next interaction rather than starting fresh.
Stateless: A system that treats each session as independent, with no memory of prior interactions. Most AI chat interfaces are stateless by default.
Zone of proximal development: An educational concept describing the range of tasks a learner can handle with support but not yet independently — the sweet spot where challenge drives growth without causing overwhelm.
Open weight models: AI models whose weights are publicly released, allowing anyone to download and run them locally rather than accessing them only through a paid API.
Distilled model: A smaller AI model trained to mimic the behavior of a much larger frontier model, making it feasible to run on consumer hardware with limited VRAM.
Ollama: A tool for downloading and running open-weight language models locally on your own machine, exposing an OpenAI-compatible API on localhost.

Resources

Things they pointed at.

00:10tool/teach skill by Matt Pocock ↗

10:00linkMatt Pocock's explanatory video ↗

00:12toolClaude Desktop ↗

00:12toolClaude Code ↗

03:50toolOllama ↗

12:50linkAI Automation Nexus (free Skool community) ↗

00:00toolHyperFrames by Heygen ↗

00:00toolGlaido ↗

Quotables

Lines you could clip.

00:59

“A stateless tool completely forgets you. A stateful tool actually teaches you and remembers you.”

Clean, punchy contrast — no setup needed, lands in one sentence.→ TikTok hook↗ Tweet quote

12:25

“It's a tutor, not a cheat code. It makes learning way more fun, but it does not do the learning for you.”

Honest reframe of an AI tool — stands out against hype content, earns trust.→ IG reel cold open↗ Tweet quote

12:45

“You can outsource all the research with AI, but you can never outsource your understanding of AI.”

Quotable maxim, no context required.→ newsletter pull-quote↗ 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.

analogy

00:00What you're looking at is a full lesson that my AI agent built for me, and it's teaching me something that I've wanted to learn for a very long time. This was all able to happen from a free skill called slash teach, and it was created by Matt Pocock. You may have heard of Matt Pocock.

00:13He's created skills like grill me and grill me with docs, and I'm a huge fan of his work. And what this slash teach skill does is it literally turns your Cloud Code agent into a personal tutor that can teach you anything. It can teach you about code.

00:24It can teach you about new framework and even about open source models, which is exactly what I'm using it for. And I don't say this lightly. It might be the most useful skill that I've installed this year.

00:33So in today's video, I'm gonna show you the actual lessons that it builds out for you, the one teaching idea that makes it work really well, and exactly how to install it and start learning in about five minutes. So get comfortable, grab some coffee, and let's hop straight into the video. Okay.

00:46So before we touch anything and get into any examples, let me explain to you the one idea that makes this whole entire concept actually work. Typically, when someone wants to learn something using AI, they'll open up their AI, whether that's Cloud Chat, ChatGPT, Cloud Code, and say, hey.

01:00Explain this to me. And without a shadow of a doubt, you'll get an answer. But that answer is always a one off.

01:05And if you were to ask that same question again, the answer would be completely different. It kinda feels like having a substitute teacher that's met you for the first time, and then the next day you get another substitute teacher, and you kinda have to reexplain everything once you meet that teacher. They don't know what you actually understand.

01:18They don't know what you're stuck on. And as soon as the bell rings, they completely forget about you. Every single time you come back and start a new session, you're starting directly from scratch.

01:25Or you're so far in your context window that you're not even getting the right information anymore because your agent's starting to hallucinate. This slash teach skill is completely different because it's stateful. So it actually knows where you're stuck.

01:35It actually knows everything that you already have learned, logs questions that you've asked so that you don't have to ask them again or you can refer to them. And so instead of having a substitute who forgets you every single time, you have a teacher that remembers you and knows exactly where you're at. What's really cool is it's also forward thinking, and it picks the next lesson on purpose directly on the questions that you're asking or where you think you are in that learning process.

01:56So a stateless tool completely forgets you. A stateful tool actually teaches you and remembers you. And that's exactly why this slash teach skill works so well, and I've been having a ton of fun with it.

02:06And it's not remembering you in some vague way. It actually logs pretty much everything you're doing once you run this slash teach command. So if you're looking inside of my operating system here, you can see I have a tab that says learning open source models.

02:17In there, I have learning records from questions that I've asked, places that I've got stuck. I've got all the lesson plans that we've created. We're about five lessons so far.

02:25And each of these lessons build on the questions and kind of where I'm at in the process of learning about open source models. It also creates a glossary whenever I ask about a certain word and what that means. And then as the lessons grow, this glossary grows along with it.

02:38What's incredibly important is that it creates a mission statement because why am I actually trying to learn this? And in my case, for open source models, what happened with Claude Fable really didn't sit well with me. And if you have no idea with what happened, you can click this video right here and go take a look at that.

02:51But that's exactly why I'm trying to learn more about open source models. It also has a notes section about everything about kinda me and how to operate and what kind of software I'm using, my hardware, how to actually operate with me.

03:02And then while it actually builds the lesson plans, it documents every single resource within this resources dot md so I can actually go reference those and really understand what's happening each lesson. So, ultimately, whenever I wanna pick up exactly from where I left off, I'm able to do that now.

03:16And so I've mentioned a few times already that I'm learning about open source models, so let's go ahead and dive into what some of these lesson plans actually look like. As you got a quick peek of in the intro, this is one of the actual lessons that my AI agent built for me using this slash teach skill. So the first one was all about open weight models, being able to run your first model on your own machine, created a beautiful HTML file that's really easy to follow.

03:36Gave me all the install commands when I wanted to install Olama, helped me understand why I'm installing Olama, and then putting in small models in there based on what kind of hardware I'm using. So it really helped me quickly pick up this simple concept of Olama and how to actually get started with open source models.

03:51What I really like is it gives me all the commands. It gives me a simple checklist that I can check off as I'm learning, and then it gives me a self check to make sure I'm actually learning as I'm doing these things. And what I really like is that it makes you take action.

04:02And there's no better way to learning than actually taking action and then answering questions about what you did, why you did it, and those simple things. Those create a very easy learning loop, at least for myself. So after every single lesson, I would go ahead and answer these questions.

04:14It lets me know that it is my teacher and not just the doc, so I could ask it all these questions. It tells me what to actually let it know before getting to the next lesson. And then at the bottom here, you can see all these sources that it provides for me with Olama docs, olama.com/library.

04:28And then the lesson one is tied to the mission.md, and all the terms are inside of that glossary HTML that you guys saw earlier, which I just realized I didn't actually show you guys, but this is what the glossary looks like. So we have our open weights model glossary, and these are all the words that have been inside of the five lesson plans that I have, all with definitions so I can kind of reference this.

04:46And when I see the word, I can go back and see, okay. That's what it means, and that's the context of the sentence. And this continues to grow as you get more and more lessons.

04:53Just scrolling through a couple of more examples. This is what lesson two looks like. This one's really cool because it actually gives me some math to actually understand what kind of models I can run on my hardware.

05:03Have I a computer from 2019, so I can't run anything crazy. I only have eight gigabytes of RAM, so I gotta be very flexible with which models I choose. This helps me understand better how to actually choose which model.

05:13And what's really cool is in these HTMLs, they're incredibly interactive, and that's why it makes learning so much more fun than my previous experiences. So I could actually type in here numbers and then get answers of everything like that, and it it just goes with the entire flow of my learning process.

05:28And, again, it gives me immediate action to take, and then I have to do a self check right after that. What's really fun too is if I'm stuck on a certain topic and I'm not just not quite understanding it, I'll just have it generate a HTML quiz for me. And I could just go in, quiz, ask questions about the quiz, ask questions about really anything related to that lesson plan, and it's gonna create a dynamic way for me to learn that.

05:48And I mentioned previously that it documents all your learning records. And based on the questions that you ask or the situations that you're in within that learning lesson, you'll have everything documented for future lessons. So in this case, I was verifying that my machine was eight gigabytes of VRAM, whatever my hardware was, so I can decide what models that I can comfortably run on this computer.

06:08And in this case scenario, I was just asking, like, what is the best open source model that I can run on this computer? I specifically asked it if I could use DeepSeek's best model, and it told me that I was out of my mind because I have nowhere near the hardware to actually run that model. It then brought me down a loop of explaining what distilled models are, frontier models, and how distilled models are just lowered versions of these frontier models that run on this incredible hardware and how you're actually able to run variations of these open source models on whatever hardware you have.

06:35But the fact that it's tracking my real progress and gives me a good standing point of exactly where I'm at in learning about open source models gives me a very comfortable feeling. But what's the thing that actually makes this thing really work? This simple philosophy is what separates SlashTeach from every other teaching AI tool that I've tried out.

06:51Matt Polkark built this on his last ten years of experience teaching people. So it's built on how people actually learn and not just spitting out random information. There are three levels to this philosophy.

07:01There's knowledge, skills, and wisdom. The first thing it does is it pulls knowledge, but not from any random sources.

07:07It actually goes to high trusted sources and vets them before embedding them into your lesson plan. Then it helps you build real practice by, again, like I said before, taking action and not just reading. And then for the last part, wisdom, it actually points you towards communities to help you surround yourself with more people learning the same topic.

07:23So if I hop over to my glossary here at the bottom, you can see that it's pointing me to the Reddit page for Olama. Because as people say all the time, you can outsource your research and all that stuff to AI, but you can never really outsource your understanding. So getting into these communities and learning from people that have actually been in your place is incredibly important, and I love how this slash teach skill implements this.

07:42But this right here is a concept that fascinated me. We all know that we all learn things differently. Like, when I was in school, I literally could not sit still, and if I wasn't interested in a topic, I just didn't care about it.

07:52But every lesson that this skill builds is pitched right at your level. So they're not so easy that you're bored, and they're not so hard that you wanna quit. They're pitched for exactly where you're at.

08:01Right in that sweet spot where you feel a little bit challenged, but you still really wanna do it, and it gets fun. And it's able to do that because, again, it's tracking your record. Based on the questions that you're asking, it's logging all that information in those lesson records that I showed you earlier.

08:13So every single lesson following that is calibrated to exactly where you're at in the current one. And teachers actually have a name for this concept, and it's called the zone of proximal development. And if you really wanna feel like you can learn anything and just go to any topic and easily pick it up, this is a concept definitely worth knowing about.

08:28Because most of the time, any other learning tool that you use will either just dump all the information on you or baby you. This one feels like it hits the right spot every time, and then it really does something you wouldn't expect. Once it feels like you've mastered or learned these skills well enough, it'll just pass you on to a community and actually step out of the way, which for me, learning about open source models is actually really exciting because I definitely wanna surround myself with some people that actually know a lot about it.

08:51And I love this because it's not trying to keep you dependent on the model to keep learning. It's trying to make you good enough at the topic so that you don't need it anymore because we don't wanna become reliant on AI. And so I know at this point, you're probably eager to get this set up for yourself, so let's get exactly into that next.

09:05If you've been following me for a long time, you know exactly how I'm gonna install this. And all I'm gonna do is hop over to this slash skills repo, grab the slash teach skill, and then hop right over to Claude code, say, please look at this skill and install it into our project. Give it the GitHub link, and then it's going to go ahead and install that skill for us.

09:22Now I already did it, so I'm not gonna enter in this prompt, but I wanted to show you guys on the screen so you can do it for yourself. And so now let's actually run this skill. So I'm gonna go ahead and run slash teach on how to actually build LLMs.

09:32Then And on the left hand side here, we only have one folder right now, which is open source models, and we should get another one once this whole process is done. And so right away, it begins setting up the entire teaching workspace for this. And first, it checks what actually exists already within the entire code base.

09:46And since learning about how to build LLMs is such a vast topic, it's gonna actually ask me some concise questions to really narrow down where we want to start learning. The first question is the why, which is gonna help us develop that mission statement. So what's the real driver behind learning how to build LLMs?

10:01And that's gonna be first principle mastery for me. Then the depth is gonna say, how hands on do you wanna get? I'm gonna say hands on code.

10:08I do really wanna learn how to actually build these for myself. And now slowly, we're gonna begin building out this entire workspace. And so it locked that plan in, and now it's immediately going and doing some web search and getting us some initial information.

10:18And a quick note while you're running this, I'm currently running this on Opus 4.8 with high effort. Matt Pocock, in his explanatory video, which will be linked in the description, ran it on medium effort. So that's also a good default if you don't wanna be spending that many tokens.

10:30If you wanted to go a little bit faster and maybe do a little bit better research, then you could just run this on the high or max effort. And I ran into a little bug, which I'm glad that actually happened, because instead of making its own workspace inside of that learning folder that I had earlier, it went ahead into the skill folder and created it inside of that, which is something I don't want.

10:48So just make sure when you're prompting this inside of an operating system and if you're not in a fresh folder, that you actually are very meticulous with where you're placing these folders. I'm gonna wait for everything to finish, and then I'm gonna show you how I merge it over back to the learning section. And so the workspace was created and actually created the first lesson for us.

11:05So this first lesson is an LLM is a next token predictor explaining us what an LLM is. We can see the four moves and how it works, the code behind it. It then gives me your turn to predict then run, which is really cool.

11:17Check for yourself. I could open these up and answer all the questions and all these sources at the bottom. But I wanna go ahead and merge this into its own workspace instead of that learning subfolder, so I'm gonna go ahead and say that.

11:27Please move this entire workspace into the learnings folder. I don't want it in the skills folder where it's currently at. So merge this entire workspace in a subfolder under the learning folder called how to build LLMs.

11:40Then I'm gonna go ahead, send that off. And now that entire thing that it just built will merge into that learning folders, which is exactly where I want it to be. So if you come across this, this is a simple fix, but just make sure that you guys are pretty meticulous with where you're actually asking these files to be created.

11:53And that's much better. Now in the learning folders, we have two subfolders called how to build LLMs and open source models. So that's how you guys can fix that really quickly.

12:01And then it's as easy as just continuing every lesson. If you're in one session, you can continue on just learning, going through the lesson plan, asking questions. Everything will get tracked, and then you'll be able to start exactly where you left off from in every new session.

12:14All you'll need to do is point your agent to the specific workspace, and you'll be good to go. And, of course, the reality is this isn't magic at all.

12:21And it's definitely not a shortcut because at the end of the day, you're learning, and you still gotta do the reps. You guys saw yourself, I'm only five lessons into my open source model journey, and I'm gonna keep going through it lesson by lesson until I feel I'm at a good spot to get pushed to a community. And that's the entire point of this skill.

12:35It's a tutor, not a cheat code, and it makes learning way more fun, but it does not do the learning for you. And I'll say it again. You can outsource all the research with AI, but you can never outsource your understanding of AI.

12:46And if any part of this video confused you at all, don't worry because I got you. I created an entire free resource guide with all the install commands, some prompts in there for you guys, and everything you need to follow alongside this video. All you'll need to do is click the link down below, enter in your email, and join my free school community AI automation nexus.

13:02You'll get access to all the resources from this video and every single resource that I've ever created. Really hope to see some of you guys in there soon. As always, if you've got any value from this video, drop a like, hit that subscribe button.

13:12It definitely helps me out a ton. Leave a comment down below on the first thing you're gonna learn using this new slash teach skill. I drop videos every Tuesday and Friday, but I'm gonna be dropping a video every single day until July 11.

13:23I really have fun making content for you guys. I'm definitely enjoying the process. And if you've made it this far in the video, thank you so much, and I'll see you guys in the next one.

13:31God bless, guys.

13:35Hey. If you're still here, that means you've made it to the end of the video. So first of all, thank you so much.

13:39Something I like to do at the end of all my videos is pray for you guys, my audience. I do this because I genuinely believe in the power of prayer. It's helped me get some out of some dark places in my life, and I can only extend that belief to you guys.

13:51And I really do love you guys and appreciate the fact that you guys watch my videos every single day and actually enjoy my content. If this makes you feel uncomfortable at all, please feel free to click off the video. I'm not here to force anything on you, but just know that Jesus loves you.

14:04But I'm gonna go ahead and pray for you guys. Heavenly father, thank you so much for this day. Thank you so much for giving us the breath of life today to chase opportunity and become better versions of ourselves every single day and glorify you.

14:17Lord, I just wanna pray a blessing over my audience, a blessing of wisdom. Lord, just continue to have them come to you and ask you questions and ask to receive wisdom because you will give it to them. And just continue, Lord, to put it in their hearts to serve their neighbor, love their neighbor, love their families, provide for their children, and then just be good people with good hearts.

14:37Lord, in everything that we do, may we glorify you and honor you. Lord, we thank you for your sovereignty, your peace, and your love.

14:44And, Lord, just continue to make us think and be where our feet are. Lord, don't let us make plans for the future. Let us be present in every single moment of every single day and just continue to better our lives.

14:55Lord, we thank you for everything. Lord, we thank you for your son, Jesus, and his sacrifice on the cross, and we just continue to praise you, love you, and glorify you. And it's in your name that we pray.

15:05Amen. Now go have a great day, guys. I'll see you tomorrow.

The Hook

The bait, then the rug-pull.

Kacper opens mid-lesson — a polished HTML file filling the screen, checklist and calculator already live — before explaining what built it. The promise in the title is not a hyperbole pitch; it is a product demo that starts on the output.

Frameworks

Named ideas worth stealing.

06:50model

Knowledge to Skills to Wisdom

Knowledge — pull from vetted, high-trust sources
Skills — build real practice by taking action
Wisdom — join a community of practitioners and step away from the AI

The three-stage learning ladder Matt Pocock built /teach around. The tool is designed to move you through all three, not just dump information.

Steal forAny teaching product or course outline — works as a positioning frame for why your approach is deeper than raw content delivery.

07:50concept

Zone of Proximal Development

Each lesson is calibrated to be just challenging enough that you grow but do not quit. The skill tracks your learning records to set the next lesson at the right difficulty level automatically.

Steal forProduct positioning for any adaptive learning tool, coaching program, or curriculum that claims to meet learners where they are.

CTA Breakdown