Big Idea

The argument in one line.

Installing Andrej Karpathy's CLAUDE.md file fixes four critical flaws in Claude's reasoning—forcing it to think before acting, minimize code complexity, make surgical edits instead of broad rewrites, and optimize for goals rather than just task completion.

Who This Is For

Read if. Skip if.

READ IF YOU ARE…

You use Claude or Claude Code regularly for coding tasks and notice your outputs are bloated, unfocused, or require constant rewrites.
A developer or AI power user who wants to improve prompt quality without switching tools or learning a new workflow.
You're frustrated that AI generates solutions that work but ignore your actual end goal, forcing you to redirect mid-project.

SKIP IF…

You don't use Claude, Claude Code, or similar LLM interfaces — this is optimization for existing Claude users only.
You're looking for general AI strategy or business implementation advice; this is a technical deep-dive on prompt engineering with one specific file.

TL;DR

The full version, fast.

A single CLAUDE.md configuration file, inspired by Andrej Karpathy's viral critique of current LLM behavior, dramatically improves the quality and precision of Claude Code output. The file installs four upgraded operating rules: think before acting and ask clarifying questions instead of sprinting on assumptions, default to minimum viable code instead of bloated implementations, make surgical edits that touch only what the request requires, and execute against verifiable goal criteria rather than working for work's sake. Side-by-side demos show the configured version asking four scoping questions before building a lead magnet, producing the same landing page in half the lines, changing only the requested button color, and looping until tests actually pass. Install it once and your foundation stays clean as you scale.

Members feature

Chat with this breakdown.

Modern Creator members can chat with any breakdown — ask for the hook, quote a framework, find the exact transcript moment. Unlocks at T2: refer 3 friends + add your own API key.

Create a free account →

Chapters

Where the time goes.

00:00 – 00:43

01 · Cold open — authority hook

Establishes Karpathy's credibility (OpenAI cofounder, Tesla AI head), teases the 10x claim and the slopocalypse frame, asks for a like.

00:44 – 02:47

02 · The 4 problems Karpathy named

Walks through the viral tweet: doesn't think before working, overcomplicates output, can't make thin-slice edits, works for work's sake (not goal-driven). Boris Cherny (Claude Code creator) and Elon Musk cameos.

02:48 – 04:03

03 · Install in one command

Shows the GitHub repo (131K stars), explains how to paste the URL into Claude Code and let it pull the file down automatically.

04:04 – 06:09

04 · Demo 1 — Think First

Side-by-side terminal: without the file Claude builds a lead magnet immediately on a vague prompt; with the file installed it stops and asks 4 clarifying questions before touching a line of code.

06:10 – 07:28

05 · Demo 2 — Simplicity (minimum code)

Same lead magnet output compared: 212 lines without the file vs 108 lines with. Pixel-identical result, half the codebase.

07:29 – 08:55

06 · Demo 3 — Surgical changes

Live edit: make the button orange with file installed changes only the button. Without the file, make the button green turns the whole site green. Comic moment, lesson lands cleanly.

08:56 – 09:59

07 · Demo 4 — Goal-driven execution + CTA

Explains goal-driven looping: Claude defines success criteria and won't stop until it passes. Bridges to a follow-up Karpathy second-brain video and asks for subscribe.

Atomic Insights

Lines worth screenshotting.

A single CLAUDE.md file inspired by Andrej Karpathy's tweet has earned 131,000 GitHub stars — making it one of the most-starred files of its kind.
Karpathy's tweet identifying four LLM problems reached 7.7 million views despite being highly technical — the appetite for this information is massive.
Claude Code's own creator Boris Cherni replied to Karpathy's critique acknowledging all four points and committing to fix them — the problems are confirmed at the source.
Current AI works to work, not to a goal — it generates output without staying anchored to the outcome you actually need, costing hours every week.
AI overcomplications are systematic: Claude will write 500 lines of code when 100 would solve the same problem without a CLAUDE.md that enforces simplicity.
Surgical editing is fundamentally hard for current LLMs — ask for a small change and you risk losing the whole file because the model can't make precise in-place modifications by default.
A CLAUDE.md file installs in one terminal command and changes how Claude plans, writes, edits, and stays anchored to your goal without any manual intervention per session.
AI trained on the internet inherits average internet behavior — average planning, average code quality, average goal-directedness — unless you override it with explicit instructions.
The before/after gap from installing a strict CLAUDE.md is not marginal — it eliminates assumptions, halves code length, and produces surgical edits instead of full rewrites.

Takeaway

Steal the format and the file.

Builder playbook

One viral tweet, one GitHub file, one install command — that's the whole tutorial arc, and it works because the demo pays every claim in real time.

Install the actual file: tell Claude Code to install this file and paste the GitHub URL. One line. Done.
Use the 4-failure-modes frame in your own content — it's already viral and maps cleanly to any AI is making your work worse hook.
Run side-by-side terminal demos — before/after with the same prompt is more persuasive than any explanation.
The 212 lines vs 108 lines cut is a format you can lift for any tool comparison: same output, fewer resources.
slopocalypse is a coinage worth borrowing — it names a real frustration in one word.
Structure: authority borrow (Karpathy) > problem diagnosis > one-command fix > live proof > sequel hook. A replicable longform skeleton.

Glossary

Terms worth knowing.

CLAUDE.md: A configuration file that Claude Code reads at the start of every session, containing persistent instructions, project rules, and context that shape how the AI behaves throughout the work — without the user having to re-explain them each time.
Karpathy file: A CLAUDE.md template popularized by Andrej Karpathy's public recommendations for addressing common failure modes in AI coding assistants — including code bloat, silent assumptions, and scope creep.
Slopocalypse: Informal term for the pattern where AI-generated code or content is verbose, over-engineered, full of unnecessary additions, and fails to do exactly what was asked — and nothing more.
LLM (Large Language Model): A deep learning model trained on large amounts of text that can generate, summarize, translate, and reason about language. Examples include Claude, GPT-4, and Gemini.
Surgical edit: A code change that modifies only the exact lines necessary to accomplish a task, without touching unrelated parts of the codebase — the opposite of a broad, sweeping refactor.
Andrej Karpathy: A prominent AI researcher and educator who co-founded OpenAI and led AI at Tesla. Known for public writing on practical AI development, including widely-shared guidance on using AI coding tools effectively.

Resources Mentioned

Things they pointed at.

02:37linkKarpathy-Inspired Claude Code Guidelines (GitHub) ↗

09:40channelDream Labs AI — Karpathy Second Brain follow-up video

Quotables

Lines you could clip.

00:35

“We're gonna put it to the test to see if we can transcend what Andre Kapathi is calling the slopocalypse.”

Coinable term — slopocalypse is a hook word that stops the scroll→ TikTok hook↗ Tweet quote

05:30

“Before I build anything per section three of the rule book, ask first on stack decisions... I need four things from you.”

Concrete before/after AI behavior — shows the file working in real time→ IG reel cold open↗ Tweet quote

06:03

“You're going to have masterpieces, not slop.”

Sharp, quotable contrast — single sentence that lands the entire argument→ Newsletter pull-quote↗ Tweet quote

06:55

“It is 212 lines of code where the Kapathi version mocked up the exact same page... about half the amount of coding lines.”

Hard number proof — specific and verifiable→ TikTok hook↗ Tweet quote

The Script

Word for word.

metaphoranalogystory

00:00Andre Kapathi cofounded OpenAI, was the head of AI at Tesla, and even had the creator of Claude Code thanking him for feedback on his AI model. So when a man of this caliber makes a tweet that gets seen by 7,700,000 people about how he thinks we're all using Claude code wrong, I'm all ears.

00:19And according to Andre Kapathi, you can 10 x the quality of your output from Claude and leapfrog your competitors simply by installing this one file. So in this video, I'm gonna show you exactly what this file is, how do we get it installed, and then we're gonna put it to the test to see if we can transcend what Andre Kapathi is calling the slopocalypse.

00:39All I ask in return is you hit that like button below, and let's jump in. So it all started here from Andre Kapathi's official Twitter account where he put out, no other way to describe this, a long ass tweet that's extremely technical talking about the four main problems that he sees with current LLMs like Claude.

00:59Now I'm still shocked myself that these things get 7,744,000 views on a highly technical tweet like this, but it just shows the amount of truth that is in here and the amount of need for something like this.

01:12Now, normally, if we're not in the AI world, we'd go through this whole post together and then try to tweak whatever processes we're doing in our businesses to fit it. However, welcome to 2026. Welcome to the modern world where someone can take that Twitter file and turn it into what's called a Claude MD file and upload it to the Internet so we can all download it in one line of code, and I'm gonna show you exactly how to do that, and then have our Claude codes operate in line with everything that Andre Kapathi says would make it better.

01:41And so this file is currently hosted on something called GitHub, and you can see it's had a 131,000 stars, which if you have no idea what stars are, it's one of the most popular files on this website.

01:53Now in this really long tweet, Andre Kapathi identified four main problems with the current setup of LLMs like Claude or Claude Code. It doesn't think before working. It doesn't plan out what it wanted to do.

02:04It just gets started, it might be running in the wrong direction. It overcomplicates the output.

02:09Lines of code, maybe 500 lines of code when it could have created what it created in a 100 lines. It's hard to thin slice edits. So you see AI these days making full, uh, five minute movies and you're like, oh, can we tweak at the three minute mark this guy smile to make him smile a lot bigger?

02:24And all of a sudden, it deletes the whole file and the whole thing's corrupted because it's really hard to sniper, pinpoint, edit, whether that be in a video, whether it be on a website, or whatever it is that you're creating with AI. And then finally, this is my favorite one right here.

02:37At the moment, AI is work driven. It's not goal driven in its execution. It works to work, not working to a goal that you may have for it.

02:46And Andre Kapathi says that this is costing you hours every week, and it's actually costing you a lot more than that in the output of your AI agent. And as I said, the actual creator of Called Code, Boris Cherni, replied to this long ass tweet and said that all these points resonate, and he's actually going to be looking to try and fix those issues himself.

03:05And even Elon Musk responded, sums up the zeitgeist because AI has been trained on the entire Internet. It's going to have basically average IQ, average creativity, average everything because everything is regressing to the mean because there's too much crappy training data in there.

03:20Now the quickest way to get this file installed so we can start to test it for ourselves is to come over to the GitHub. You can copy paste the actual URL, uh, or you can come down to this thing here, it'll give you the URL to clone and copy it. And this is the awesomest thing about AI, is you go to your Claude code.

03:37So whether you use that in terminal like I do or maybe use it in the app, it doesn't matter. All you have to do is say, please install this file for me, and then paste the URL.

03:48And it's gonna go ahead and pull that entire MD file down onto your local device so it can start operating through Andre Kapathi's upgraded skill set. So now that it's installed, let me show you the difference that this makes to you operating your business.

04:01We're gonna go through the categories that we went through of the four fundamental upgrades that Andre Kapathi is recommending that your LLM makes. Starting with thinking first.

04:11When you give Claude Code some instructions, it's gonna start sprinting unless you have this file installed. This is gonna make sure that it sprints in the right direction. It's gonna not assume things.

04:21It's gonna make sure it never hides confusion, and it doesn't surface trade offs.

04:25It's even gonna push back when warranted and stop when confused. It's not gonna build for building's sake, and it's gonna ask you more questions. Okay.

04:34So I've pulled up two versions of terminal, which, again, if you don't use it in terminal, it's gonna look codey in advance. Don't worry about it. It's just the exact same as running inside the app, but you have way more features.

04:43I actually have a full video on this and showing you how simple it actually is. But I've typed the same prompt into both of these terminal windows, one with the Kapathi MD file installed and one without. So I said, can you make me a lead magnet for Dreamlab?

04:57So something that I can put into this video and then get an email in return for and add people to my email list. The first instance, which does not have the Kapathi MD file installed, says, I'll check the Dreamlabs context first then build the lead magnet. It went ahead and just literally built the lead magnet based on this very, very vague prompt.

05:15However, it's not bad because it did link straight into my active campaign and straight into the right email folder and then hosted it on the correct website. Once again, you wanna see exactly my setup, I have a video breaking that down, but it did make a lot of assumptions. On the second instance of terminal where this Kapathi file is installed, I said, can you make me a lead magnet for Dreamlabs?

05:36He says, before I build anything per section three of the rule book, ask first on stack decisions and section 7.1, think before acting, I need four things from you. What hook do you want?

05:46What format do you want? What list do we add them to? And what URL slug, where does it host it?

05:51Once I have those four, I'll draft the page in the standard dark aesthetic single file HTML, show it to you locally, and I'll only deploy it once you give me the green light, of which, of course, is going to lead to way higher quality outputs. But if more context and high quality outputs isn't enough because initially you're like, wait, that's just more work on my behalf.

06:11Yes. Of course, is, but you're going to have masterpieces, not slop.

06:15The second part, however, is even stronger than that. It's simplicity, which is basically minimum code. Everything's built from code, whether it's videos, photos, websites, lead magnets, PDFs, it's all going to be code.

06:26We are training our core to be as simple as possible. So as we continue to build on top of it, its foundations are clear and clean. And so when I asked the terminal window that does not have this Kapathi file installed for a lead magnet, this is what it gave me.

06:40Get my exact AI command center, full install config, hands free voice control on Mac, every skill and agent, an MTP I run, how to point it at your own business, which is pretty good because I have given it a lot of context on my business. Send me the setup. Put your email in here, and it's all hosted on my actual website.

06:55It's not bad at all. But it is 212 lines of code where the Kapathi version of it mocked up the exact same page.

07:04So this is the old one, and then this is the Kapathi version. Literally no difference. It's almost pixel perfect.

07:10But it has about half the amount of coding lines that go into it, which is pretty impressive, especially as you scale these things and load time you want to be as quick as possible, and then building on top of it, you want that foundation to be as clean as possible. And we get all this efficiency just by downloading that one file.

07:25Now the third skill we get in this beautiful skill pack is surgical changes. So it wants Claude to touch only what it must touch.

07:34Clean up its own mess after creating it. And when editing existing code, don't improve adjacent code, comments, or formatting. Don't refactor things that aren't broken.

07:44Match existing style even if you do it differently. And if you notice unrelated dead code, mention it. Don't delete it.

07:50Because Claude is historically bad. If you ask for the color of this button to be changed, it might tamper with, say, this box up here and all of a sudden the title is also a different color. But with this file, it's going to have precise edits.

08:02And I mean, let's run the test together. Let's see if we can get an orange button here. So I'm just gonna paste that URL right here.

08:07I'm gonna say, make the button orange, please. Play, should I say.

08:15And he says done. The button is now an orange gradient. I'm gonna refresh this page and it is orange and nothing else has changed.

08:22So now let's let go back and test the original one. I'm gonna copy the URL. I'm gonna come back to our original terminal here without the Kipathi file installed.

08:30I'm gonna say, can you please make the button green? Okay. So it says it's changed the button.

08:38I'm gonna come back and refresh it, and the whole site is green. That's obviously a joke, but I'm sure you've had your Claude code just touching things that it shouldn't be touching.

08:49Now the last one, which is my favorite upgrade to my Claude since installing this file, is that makes every action goal driven. You get to actually define success criteria for every task or project that you're giving your AI, and it's gonna loop and continue to work until it's verified, hit that goal, and become successful in its effort.

09:11So it's going to basically give itself test results and make sure that it's hitting the a plus in the criteria that you gave it. This is when you see those jokes where it's like make me a $100,000,000 business, make no mistakes, keep working until we have a 100,000,000 in the bank.

09:24And so for example, if you have something that has a bug in it, you can come in and write something like fix the bug, and it's gonna transform that input into write a test that reproduces the bug so we know exactly how we got it, then make it pass. It will not stop working until it actually passes that criteria. The best example of this is actually in my other Andre Kapathi video where I show you exactly step by step how to set up Andre Kapathi's second brain for business that is centered around your business goals and needs.

09:52I'll pop it up here on the screen somewhere, so go and watch that. And if you haven't, hit that subscribe button below. Thanks so much.

09:57I'll see you in the next video.

The Hook

The bait, then the rug-pull.

When a tweet from Andrej Karpathy hits 7.7 million views — and the literal creator of Claude Code replies saying all these points resonate — you stop and pay attention. Dream Labs AI did the work so you don't have to: one CLAUDE.md file, one install command, and suddenly your AI stops sprinting in the wrong direction.

Frameworks