21:39Dream Labs AI · Tutorial
This "Karpathy System" could 701x your AI Workflows
How to point Karpathy's open-source AutoResearch at your cold emails, ads, and website so AI runs experiments while you sleep.
June 9thThe argument in one line.
A single line telling Claude to log every correction as a one-line rule under a Lessons heading in CLAUDE.md turns a static preferences file into a compounding memory that prevents repeat mistakes across every future session.
Read if. Skip if.
READ IF YOU ARE…
- You already use Claude Code or a similar AI coding assistant regularly and want your corrections to actually stick between sessions.
- You maintain a CLAUDE.md or similar persistent-context file and keep re-explaining the same preferences every session.
- You're building AI-assisted workflows (agents, automations) and want a low-effort way to make them improve over time without manual file editing.
SKIP IF…
- You don't use an AI coding tool with a persistent memory/context file (CLAUDE.md, .cursorrules, etc.) - the technique doesn't apply.
- You're looking for a deep technical explanation of how Claude Code's context loading works - this is a practical demo, not an internals breakdown.
The full version, fast.
Most Claude Code users either run with no persistent context file (get generic, unpersonalized output every session) or have a static CLAUDE.md (get personalized output, but corrections vanish once the session ends). Boris Cherny's fix is one added instruction: when the user corrects Claude or Claude catches its own mistake, before continuing, append the lesson as a one-line rule under a '## Lessons' heading in CLAUDE.md. The host proves it with a live test - three Claude Code sessions build the same sales page from an identical prompt. The beginner (no file) produces a generic, unbranded page. The intermediate (static CLAUDE.md) produces on-brand copy and imagery but ships unwanted design choices (green accents, rounded corners). The advanced session gets verbally corrected once ('always light mode, no rounded corners, no green'), which Claude writes into its own Lessons section - and a completely fresh session, given only a one-line prompt with no restated rules, correctly outputs light-mode, square-corner, green-free pages, even auto-correcting a green icon it initially shipped by cross-referencing its own lesson file.
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 – 00:40
01 · Cold open
Hook direct to camera promising Boris Cherny's magic snippet.
00:40 – 01:29
02 · Boris Cherny's January post
References an older Cherny Twitter post on his Claude Code workflow; most of it is stale except the CLAUDE.md habit.
01:29 – 02:40
03 · Three levels of Claude Code user
Beginner (no file) vs Intermediate (static CLAUDE.md) vs Advanced (self-learning CLAUDE.md) framework introduced.
03:05 – 04:12
04 · The self-learning snippet revealed
Exact text of Boris Cherny's snippet shown on screen, plus a short clip of Cherny explaining the idea.
04:12 – 07:05
05 · Three-way live test setup
Three parallel Claude Code terminal sessions each given an identical prompt to build a sales page for the host's Dream Labs offer.
07:05 – 08:06
06 · Beginner results
No CLAUDE.md session produces a generic, unbranded sales page with a made-up logo.
08:06 – 09:33
07 · Intermediate results
Static CLAUDE.md session pulls in real branding and copy but ships unwanted green accents and rounded corners.
09:33 – 10:24
08 · Teaching the advanced session
Host verbally corrects design preferences; Claude writes them into its own CLAUDE.md under a Lessons heading.
10:24 – 11:30
09 · Self-learning MD file in action
A brand-new session, given one line of prompt with no restated rules, correctly applies the logged lessons - including auto-correcting a green icon.
11:30 – 12:53
10 · Final comparison and sign-off
Side-by-side of all three final builds; pitch for the Dream Labs paid community.
Lines worth screenshotting.
- A CLAUDE.md file alone only solves for one session at a time - corrections you give Claude vanish the moment that session ends unless you add a rule to persist them.
- The fix is one added instruction: tell Claude that when you correct it, it should write the lesson as a one-line rule under a Lessons heading in its own memory file before continuing.
- In a live three-way test, a Claude Code session with no CLAUDE.md produced a generic sales page with a made-up logo and no brand voice.
- A session with a static CLAUDE.md pulled in real branding, logo, and copy references, but still shipped design choices (green accents, rounded corners) the creator didn't want.
- After being corrected once on design preferences, the self-learning session wrote those corrections into its own CLAUDE.md as permanent rules without being asked to.
- A brand-new Claude session, given only a one-line prompt and no restated preferences, still produced light-mode, square-corner, green-free output purely by reading its own prior lessons.
- Claude auto-grayscaled an icon it had shipped in green after cross-referencing a lesson it had logged about hating the color green - a self-correction with no human intervention.
- The technique generalizes beyond code or design - any repeated mistake (tone, formatting, emoji use) can be captured the same way so it never recurs across any future project.
One added instruction turns corrections into permanent rules.
WHAT TO LEARN
A persistent context file only compounds in value if you explicitly tell the assistant to log every correction as a rule before continuing, otherwise each session starts from the same static baseline.
- A persistent context file (CLAUDE.md or similar) only personalizes output for the current session unless it's also told to log corrections permanently.
- The specific instruction that makes this work: when corrected, write the lesson as a one-line rule under a dedicated Lessons section before continuing with the task.
- In a live test, a session with no context file produced generic, unbranded output; a session with a static file produced on-brand output but still repeated unwanted design choices.
- Once corrected a single time and told to log the correction, a fresh session with zero restated instructions still applied the learned preference automatically.
- The AI can self-correct mid-task by cross-referencing its own logged lessons, such as reverting a color choice it had just made without being asked to.
- The technique isn't limited to code or design choices - tone, formatting, and any recurring mistake can be captured the same way so it stops recurring across every future project.
Terms worth knowing.
- CLAUDE.md
- A plain-text file that Claude Code reads at the start of a session, used to store a user's preferences, business context, style, and rules so the assistant doesn't start from zero every time.
- Self-learning snippet
- A specific instruction added to a CLAUDE.md file telling Claude to write down any correction it receives as a permanent one-line rule, rather than only applying the fix for the current session.
- Lessons section
- A heading inside CLAUDE.md (e.g. '## Lessons') where an AI assistant appends rules derived from past corrections, so future sessions automatically avoid repeating the same mistakes.
Things they pointed at.
00:00linkBoris Cherny's January 2026 Twitter/X post
Lines you could clip.
03:25
“When I correct you or you catch yourself making a mistake: before continuing, add the lesson as a one-line rule under Lessons, so it never happens again.”
11:47
“It literally edited the green out of an icon because of what it remembered from that MD file.”
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.
metaphor
00:00Boris Churney, the creator of Claude Code, has just revealed a tiny little magic snippet of text that you can paste to your Claude to make it smarter, faster, more personalized to you, a self learning work machine. Claude will remember every single mistake it's ever made so that it never makes it again. It'll learn your style and understand your workflow.
00:19But best of all, it'll get smarter every single session that you guys have together. So in this video, I wanna reveal Boris' magical little text snippet, show you exactly how powerful it really is, and have you copied into your own Chord so you can supercharge your workflow outputs. Hit that like button down below.
00:35Let's jump in. So you do have to be extremely careful when researching how Boris uses Chord because there is a lot of misinformation churning around the Internet.
00:46For example, this post that Boris actually posted himself on Twitter back in January 2026 shows in detail exactly how Boris uses Chord. However, most of this information is now outdated, and even Boris admits that there's only a few fundamentals that still hold true from this post. One of the fundamentals that is still holding true is this cool little tool that Boris uses called a Chord MD file, which put simply, it's basically just a text file that has all your preferences, your style, your tone, your business, your mission, all the things that you want to put in Chord's long term memory to constantly reference and personalize its results back to you you put in this file so we can reference it in every single session.
01:26But, of course, this is not the magical snippet of text. To understand how powerful the snippet of text is, I wanna show you the three levels of a Claude code user. Starting with level one, the beginner.
01:36So the beginners are the people out there raw dogging Claude. What does that mean in the industry? It means you're not using a Claude MD file.
01:44So every single new session that you start with Claude, there is no long term memory, no understanding of your business, your style, your tone, your mission, your voice, or any of that. You're just basically starting all over every time you start a new session. And therefore, beginners are gonna get very generic and poor outputs from Chord.
02:00Level two, the intermediates. So intermediate users are creating a Chord MD file like Boris. And I've already explained how this file gives your Chord a long term memory across each session and styles its answers to your preferences.
02:14But there is one massive problem with this file. All the corrections that you make with your Chord whenever it makes a mistake or it doesn't understand what you're saying or gives you the wrong response back vanish the moment that that session ends.
02:26And therefore, you might wake up in the morning, start a new session, and it's gonna make all the same mistakes all over again. , and it's not getting smarter with every conversation that you have with it. However, does Boris have the solution for you?
02:40Introducing level three, the advanced user.
02:44So the advanced users are copying Boris' magical snippet of text, which for the first time ever allows your AI, your Claude or Codex, whatever it is that you're using, and specifically your MD file to become self learning and improve with every session that you have.
03:01And that's because the snippet of text is this. It's called a self learning snippet. Boris writes, when I correct you or you catch yourself making a mistake, remember this is talking to Claude AI, before continuing, add the lesson as a one line rule under lessons in your Claude MD file so it never happens again.
03:21And then inside the MD file, Claude has a section called lessons where it's going to apply all the rules, all the mistakes, all the corrections that you've ever given to it in session to the MD file so it never makes that mistake again. This is Boris talking about it. But it's I think it's just, like, the most important idea when working on this stuff is, like, every single time Quad makes a mistake,
03:42I don't tell Quad to do it differently. I tell it to write it to the QuadMD or to, like, make a skill or or something to do it differently. And if you can do this, then Quad can just, like, run forever.
03:53Okay. So how powerful is that snippet of code? I actually wanted to run a test with you, but we have three sessions of Claude code running.
04:00One, the advanced with no Claude MD file. Two, an intermediate with an MD file that is not self learning. And three, Boris' magical snippet of code deployed inside this Claude MD file.
04:11And I'm gonna ask each one of these sessions to build me a sales page. K. So we have the beginner, intermediate, and advanced setups ready for us to test.
04:21Now I'm running them inside of my terminal on my computer. You can run them anywhere you want. Wherever you use Chord, you can interact with it and point it towards an MD file and a self learning MD file.
04:30I've got a sandbox running on the one on the left, which means there is no MD file at all. It's just RawDog and Claude as they say in the industry. The one in the middle is pointed towards my Dreamlabs folder.
04:41So you can see Dreamlabs right here. They're both running on Fable five. Fable five is the same model.
04:46And then this last one down here is running on the Dreamlabs self learning Chord MD file. And so what we're gonna do is we're gonna paste in the prompts.
04:55So I'm going to be sending this off. We're basically asking to build a sales page for my membership offer for the Dreamlabs channel. Gonna send it to the intermediate one as well.
05:04Quite a long prompt because we only like to give it as much information as possible. You can see the business I run is Dreamlabs, a YouTube channel that teaches business owners how to supercharge their AI, get 10 times better results by distilling down and reverse engineering the systems of the best builders, leaders, and AI engineers in the world.
05:20Okay? This is the offer down here. This is the price point.
05:23This is what it actually is on the inside. So it's not working in the dark. It has a decent amount of information even though this beginner one doesn't have a Chord MD file.
05:32The middle one, however, does have a Chord MD file. I haven't started the advanced one yet. We'll get to that in just a second.
05:38I wanna show you the MD file of the intermediate level. So here is Dreamlabs Chord MD, which is this file that shows Claude.
05:48It gives it a reference on everything that I like, my voice, my vocabulary, my writing rules, my content system, what we stand for, the business, the mission, everything we're anchored towards. So it becomes like a worker that knows all the context I needed to know in order to help drive my goals forward literally even while I'm asleep.
06:06The agents that are running while I'm sleeping are referencing this Claude file twenty four seven without having to reset their memory over and over again. And, hopefully, in this example, you'll see the difference in a web page. What it's also got down here are links to other MD files, such as my style guide here, where it goes and looks at what sort of style I like to build things in.
06:27And so the only thing the intermediate step is missing, however, compared to the advanced step is one line. And so this MD file is the advanced one where there is a section that says self learning. When I correct you or catch you making a mistake before continuing, add the lesson as a one line rule under hashtag hashtag lessons so it never happens again.
06:47Down here, I've reset my Claude file for you. Claude adds rules here. This is where all the lessons that Claude learns or the corrections I give it, all the mistakes that I help it understand that it's made and never wanna make them again, they will be listed down here.
07:01Okay. So let's have a look at the beginner results.
07:05This is their file right here that they made, the sales page for Dreamlabs. I'm gonna copy that. If you scroll up, you can see once again all it had was this prompt right here.
07:14And so remember, this is Fable five. It's an excellent model, but it should be pretty generic. So this is the file right here.
07:20So you can see Dreamlabs. That's not actually my logo. They've made up a random logo here, which of course, they have.
07:25They don't have any references or files because they don't have that Claude MD. But to be honest, it's not bad at all. You're not behind because you're not smart enough.
07:33You use Chateappity every day, write saves and emails, good enough summaries. It's fine. So it's not bad.
07:37It is very generic. This is not the AI. This is a system around the AI you already have, and it's very wordy, very texty.
07:43So these are some of the things that I've given it in reference on the prompt that we actually have. 11 upgrades, one clear order. Memory, context, workflows, agents.
07:52Not very visual. It's very text heavy. The design's clean.
07:55It looks nice. The frequently asked questions are looking pretty good as well. And I like the design of the button, but it is very, very generic because it had no reference files, no style guides, no anything.
08:06Alright. So the intermediate results are fascinating. First of all, because it's plugged in to my ecosystem, it could deploy it straight to my website.
08:15It could have Stripe already integrated. It could have my email list already integrated. This is why you wanna set up Chord in an excellent way like we do inside our Dreamlabs community because it makes things so much easier.
08:26This is the URL right here which we'll copy out, and you can see it's got our checklist going down the middle here. And it cooked and referenced not just the prompt, but it also referenced the MD files that we gave it. Here is the result, which there's two things to point out here.
08:43Design wise, it is kinda generic. It almost went back to the Opus 4.8 sort of design. But if we look past the design, because that's way easy to tweak, you can see it has way better copy.
08:54It has pictures and references for me. So this pulled it from my actual YouTube channel. It has the heart of what we actually offer.
09:02I would like to see more visuals because once again, this is a little bit more text heavy. But the 11 path upgrade build, that's what we got down here. I just need visuals for it.
09:09We've got our AI agents that we can deploy. You've got your mentor cartridges that you can put in and use as a council or board of directors on your business. You've got the plugins, the tools.
09:17It is slightly more visual, which is nice, and it's got the right price point, but that was also in the prompt. Scrolling down, you can see it's cropped to the background out there and put me down there. And it's got my actual Reel logo in here all because it had references in the Chord MD file.
09:33But this is where you ascend from intermediate to advanced because all the mistakes that it's made on this page, if I ever make another page in the future, if I'm making a checkout page, if I'm making a lead magnet page, anything that I wanna create, it's gonna make the same mistakes unless I update that Claude MD file. And that's why Boris Czerny automates this process.
09:52So in the advanced terminal, I'm gonna say, I always want light mode web pages, and I'm gonna send that off.
10:02I also am gonna say, I hate with a passion rounded corners and the color green.
10:09Never use them on a web page. Okay. So now that I've given that feedback, I come into the folder section on my computer, the DreamLab self learning, where the brand assets, its past builds, the Claude MD file, and the style guide sit.
10:24I open up that Claude MD file. I can expand this a little bit. And you can see right here, this is the self learning line, the magical line from Boris Czerny, and under there, it started to plot those lessons.
10:35Always build web pages in light mode, never use rounded corners on a web page, never use the color green on a web page, or as an accent highlight, wash, glow, chip, or success date. Benjie hates it, which is fantastic. So now if I open up a brand new instance of Claude and it's obviously pointed towards that MD file, I'm gonna ask it using WhisperFlow.
10:55Can you please build me a sales page and a lead magnet page and a cart checkout page? That's the only prompt I'm gonna send it, and it should remember all those rules and rebuild this in line with the new lessons. I'll show you the results once it's done.
11:11But before I show you, remember, I'm just building HTML pages. Whether you want it to use more emojis, stop using em dashes, whether you want it to address you in a certain way, whatever mistakes it makes as it makes them every time you use it, it's automatically gonna be added into its MD file so it never makes that mistake again no matter what project or what thing it's working on for you.
11:30Okay. So on our final test with the self learning MD file, Chord has got it all completed. Of course, it's plugged into my ecosystem, so it could push them live, put our Stripe in there, have them on an actual website, but we're gonna hold off to that for now.
11:43The interesting thing that happened here, the WhatsApp icon shipped green in the connectors row. So it knows I hate green, remember.
11:50Apparently, hate green. Don't actually hate green. Nothing wrong with the color green.
11:54It's a beautiful color. The whole icon row is now grayscale. It So literally edited the green out of an icon because of what it remembered from that MD file.
12:03So I've got three files here, one, two, three. I'm gonna open them up in Chrome, and here they are in the exact same style using my actual logos. No rounded corners, no green.
12:15The checkout page here, again, you gotta embed Stripe here, but there is no green, no round corners. And, of course, the sales page that they redid looks a lot stronger actually than originally did.
12:26You can see this is where there was green, so they just grayscaled the entire row. And, honestly, it did a really good job. So I'll leave Boris' text snippet in the description below.
12:36If you wanna supercharge your AI and get 10 times the results based on everybody that I've been studying for the last couple of months. Come and join us inside the Dreamlabs community. We would love to help get you your dream AI system set up so you can crush your competitors.
12:50Thanks for watching. I'll see in the next video.
The Hook
The bait, then the rug-pull.
A tiny addition to a CLAUDE.md file, credited to Claude Code's own creator, promises to turn every correction you ever give an AI coding assistant into a permanent rule - so the video puts it through a live, three-way test to see if that claim holds up.
Named ideas worth stealing.
01:29model
Three Levels of Claude Code User
- Beginner - no CLAUDE.md
- Intermediate - static CLAUDE.md
- Advanced - self-learning CLAUDE.md
A maturity ladder for how much persistent context and self-correction a Claude Code user has configured, used to structure the whole video and the live test.
Steal forframing any before/after AI-workflow comparison video
How they asked for the click.
VERBAL ASK
12:43product
“Come and join us inside the Dreamlabs community. We would love to help get you your dream AI system set up so you can crush your competitors.”
Soft-sells throughout (brand visible inside the demo pages themselves), then an explicit direct pitch in the final 15 seconds for the same $67/month membership the demo was building.
MENTIONED ON CAMERA
FROM THE DESCRIPTION
PRIMARY CTAWhere the creator wants you to go next.
Visual structure at a glance.
