Make Claude 10x Smarter For Free (Karpathy's Hack + 4 More) transcript

In today's video, I'm going to reveal five GitHub repos that fix Claude's worst habits. I don't know about you, but it seems like a week or two after a new model comes out, it just starts getting really dumb. It happened with Claude Fable. It happened again with GPT Astra. So, in today's video, I'm going to give you five hacks that can actually make Claude smart again. One of them is going to help with context. So, if you're spruled across multiple chats, it's going to help and tie all of your memory and context together.

One of them is a skill that's especially helpful if you have ADHD to stop your coding agent from burying an answer to actually give you the response that you need to see quicker. One of them is an absolute hack. It's going to help you get better writing outputs and actually humanize your text. The fourth one is going to save you a bunch of money by slashing token expenditure by up to 65%. And the last one is a system that Andre Kapathy uses.

He was in the founding team of OpenAI and then he later joined Anthropic. I'm going to show you one single file that will turn your Claude code from an idiot into an absolute monster. And the best part is all of the five repos that I'm going to show you today are available for free on GitHub. What I'm going to do is I'm going to compile them all into an MD file with the setup instructions. So all you need to do is drag and drop the file into your Claude code and it will set all of them up for you.

That's going to be available for free down below in my free school community under the resources library just like every other video here on the channel. Now the first one I want to show you is called Claude Mem. This gives your cord persistent context across sessions for every single agent so it doesn't forget stuff. How much time have you spent re-giving agents context? Even if you have an MD file, sometimes it seems like they don't read the file and every time you open a new chat or a new session, you're like reexplaining context or it's getting things wrong.

This solution is awesome because it captures everything your agent does during a session. So, if you're working on some research for a project, it's going to remember how you did that research. It's going to compress it with AI and it's going to inject relevant context that it saves from one session into the next session. So, this actually doesn't just work with Claw. This also works with OpenClaw, Codeex, Gemini, Hermes, etc.

But it's a game changer and there's a reason why it has 94,000 stars and over 162 contributors on GitHub. Okay, so it's all set up. Let me show you how it works. So, I'm going to put in some random facts about me. Obviously, these aren't real facts, but we're trying to test it out. So, my favorite color is burnt orange. My first car was a 2004 silver Toyota Corolla. and a bunch of other responses as well. Now, usually if you just open a new chat or a new terminal session in claude code, it wouldn't remember this stuff because this isn't in the same project.

But watch with the claude mem installed how it remembers. So, I'm going to open a new terminal session. I'm going to switch the folder to a folder with no data. And I'm going to ask it what my favorite color is, what my dog's name is, what my first car is, and who laugh when I got stung by a jellyfish. You can see it's searching Claude me in the tools. So, it's using the plugin. And you can see it just found all the data.

My favorite color is burnt orange. My dog's name is Biscuit. My first car was a to crawler. My friend Tom laughed when I got stung by a jellyfish. If this weren't installed, it wouldn't have remembered this cuz it's a new session in a new folder. And if you're using the terminal or VS Code, unless you have all of your memory updated into a memory folder, it's not going to remember everything. So, you might be wondering, does it remember everything?

Well, it doesn't. And that's by design because otherwise it would bloat clawed too much. Instead of storing transcripts, the worker inside the plug-in compresses the key information into short summaries with the model and it stores those summaries. So instead of the whole raw session, it just remembers the key details similar to how the claw chat will remember key details. So if you message a bunch of claw chats, you'll know it pieces things together.

But the same isn't true for coding sessions. And if you're like me and if you're utilizing tools all the time and custom plugins and repos, then you want that same effect across all of your coding sessions as well. Just keep in mind that it can slow claude down a little bit, but I think it's worth the trade-off in most cases. And it's not always calling on claude mem all the time. It's mainly doing so when it needs contextual information and then it will, you know, quickly look through the summaries.

Just the same as if you had a folder with a memory file, it would look through that summary. I would say this is best used for someone that works in claude code a lot or codeex a lot and you work for multiple days across the same project. Then it's going to fix the problem of needing to reexplain context constantly. And that's been a massive problem for me in the past. So this has been a game changer. And honestly, since installing this, I don't even think about it.

It just does it automatically. And it's one of those things where you might not even realize how bad Claude actually is at keeping memory until you install a plugin like this. All right. Now, moving on to repo number two. This is great. It's called I have ADHD, and I literally do have ADHD. So, this is something that is really helpful for me. It's essentially a skill to stop your coding agent, in this case, Claude, from burying the answer that you want.

So it's like the ADHD friendly version of using AI. We'll test it out ourselves in a second, but you can see the response from Claude before versus after using this repo. So before it uses all this language, especially the new models. Great question. Let me think about this. Your author has a few moving pieces. This, this, this. One approach would be this, this, and this. Whereas if you're developing a product or if you are trying to get a solution to a problem, usually you only want to know what your next step is.

So instead of this big bloated response, it just says run this then edit this, open, replace, run, next, paste this if the test fails. It just makes working so much easier, especially if you're coding, if you're developing, if you're building landing pages, all that stuff. It really shrinks down the response time. And essentially, the way it does this is through a skill.md file, which you can obviously see inside the folders here, which has 10 rules.

One, lead with the next action, so it tells you the next action up front instead of going with all this wordy stuff. Two, number multi-step task. Three, end with one concrete next step. Make wins visible. No preamble. all of the stuff that in the past you had to like manually inject that into your instructions file. If you just download this plugin, it'll just happen throughout your entire session automatically on the sessions where you want to use it.

So, typically where I would use this is on a coding task, anything where I'm not brainstorming or being creative. Basically, whenever I'm doing something with an agent, I'll use this instead of just using the raw clawed session. So, we'll start a new session. If you want to use it, use for/ I have ADHD. Remember, it's just a skill. So, just like you trigger any other skill, you just want to use the slash command. You don't want to use it all the time.

Obviously, if I'm doing creative writing, I don't want things compressed, but if I'm developing, I generally do. So, now it's invoking the skill. You can see that ADHD mode is on. And I'll use a simple prompt. Let's say I'm just trying to get some work done, and I'm deciding where to put my CRM. Should I do it in Notion or Google Sheets? I'm going to ask it which one it recommends. And you can see that it formats in a very specific way.

Google Sheets. Three reasons. 1 2 3. Setup time. Next. Or let's pretend we're developing something like a website. I said, "I'm working on an AI edge website. Three things are broken. I state the issue. What do I do first? Walk me through fixing it." So you can see once again it gives the output up front. It gives a list of steps and then at the end it tells us what we need to do next. This is much better, more concise than the wordy schmurdy claude without the I have ADHD skills.

So once again won't use it all the time but when I do it's when I want to just get stuff done. It's really good. Now the third repo I think is essential for people because I am sick in business of getting these AI sounding proposals. I'm sick of seeing AI tweets. I'm sick of AI language. You think your writing doesn't sound like AI? Well, it does. If you use AI enough, you can tell what's clawed. You can tell what's GPT.

There is a repo called Humanizer, and I'll show you the difference between an output before or after because it's honestly a gamecher for writing human sounding AI language, which is very important, especially if you're into writing, especially if you're creating content. You want stuff to sound like a human wrote it. You don't want to sound like it's AI. So, let's ask Claude to write me a tweet about why most people are using AI wrong.

Punchy for a founder audience under 280 characters. Most people treat AI like Google with better manners. Founders who get it treat it like a team. Context talks, standing instructions, feedback loops, ownership of the whole workflows, not one-off tasks. You don't prompt the team, you manage one. This contrast phrasing is so obviously AI. Speed is the wrong metric. Leverages like all of these contrast phrasings I hate.

Stop asking how do I do it? Start asking this. So, let's ask it now. Use #humanizer to write these three tweets again. And you'll see the difference between the AI writing and the humanizer writing. By the way, you can also train it you so you can improve upon the base skill by rewriting it to sound like your voice. So if you go for/humanizer, here's an example of my writing and then give it a bunch of stuff you've written.

Now humanize it like this. You can update the skill. So then it will sound more like you in future. So I've developed a writing skill that sounds like me to speed up time when it comes to writing stuff. And what I'll do is I'll just like voice dump using Whisper Flow if I want to like write a post or something and then it will just like reward it. It already knows my language. It already has my natural intonation. So the more you actually speak, the more you write, the more it sounds like you over time.

And you don't even need to find custom examples of your writing. You can literally just ask Claude to go back into the chats and pick up on how you speak, the vocabulary you use, and then make that uh voice match on a skill update. So these are the rewrites. As you can see, it got rid of all of the notexpert Y closes that AI just seems to love to use. It got rid of the rows of fragments, context dock, stand instructions, feedback loops.

Like these are classic AI tells. It got rid of the force triads. Like all of the classic things that AI loves to do in its writing. So for example, if AI is only saving you time, you're underusing it. The founders pulling ahead use it for work they could have never hired for or funded. The question changes from how do I do this faster to what can I try now that I couldn't before? Like it's not perfect, but it definitely sounds a lot better. and it's going to basically do that initial scan to make sure there are no of those classic idiosyncrasies that AI typically uses in its writing.

I actually really like the final phrase of this one. Most founders use AI like a fast intern. Ask, get an answer, move on. The ones getting the most from it treat it like a co-founder. They hand it the whole problem, constraints included, and let it argue back. It's slower per question and worth far more per decision. It just sounds more natural, especially the closing sentences. So, definitely worth using if you're writing.

You can inject your own style into it as I said and then you can make it like a custom writing skill essentially and you just adapt it over time. But it's a good base to start from. And if you're interested, you can go onto the GitHub and you can see exactly what it's changed. But there are 25 patterns that it corrects by using the skill. So this is the beauty of downloading these skills off the internet. People have already put in the work.

Like in this example, there's already 20 contributors that have put in the work to build this and you can just hop onto it and just make it slightly better. All right. Now for the final two, I feel like I've saved the most powerful ones until last. The first one can cut 65% of your token usage which will literally save you money. And then the last one is based on the real system that Andre Kapathy uses founding member of OpenAI and now at anthropic.

So this skill that can save you money is called caveman. Why use many token when few token do trick? Viral skill plus proxy for coding agents that cuts 60% of tokens by talking like a caveman. Your coding agent bills by the word and writes like it knows that. Caveman makes it stop. So, if you guys don't know, token expenditure is actually just based on words. So, if a normal agent gives an output like this, it'll give an output of 69 tokens.

If a caveman agent is giving an output like that in 19 tokens, then it not only saves you time having to read it, just like the I have ADHD skill, it actually saves you money because AI billing actually bills based on tokens or words. So, your agent pays for every token it writes or every token it reads. So, the more text you send an agent, the more you're being charged, the larger the outputs, the more you're being charged.

Even if it's slop, most agents write like a cover letter and read like a fire hose. Kman attacks both ends in the agent you run and in the one you build. One rule, the skill shrinks what the agent says. The proxy shrinks what the agent reads. The middleware does the same inside your own code. Started as a joke, hit 4,000 stars in a week, now passed 100,000 stars and 107.2K now. So, it's clearly gained a lot of popularity.

And it's simply because it's a really effective way to save money while saving time at the same time. So if I go forward/caveman, I can activate the skill inside claude code. Technically, it's a plugin that you download. So it's could be persistent across your sessions. So it says kbab mode full send task. So we'll show it on native text. I'll ask it what's the difference between a skill, an MCP server, and a plugin in claude code.

When would I use each one? And you can see it's simplified everything. Use skill when repeatable workflow house style review checklist project convention. Cl has the tools. Cheap, no runtime, no schema cost, use MCP. And then a rule of thumb, knowledge equals skill, capability equals MCP, distribution equals plug-in. So now I've turned off caveman mode. And I'll ask it the exact same question. And you can see that it's already much more verbose.

So if we actually run the calculations on tokens, you can see caveman used 409 tokens. Normal used almost 600 tokens. Caveman was almost half the words and almost half the characters as well. So it saved about 31%. In some cases, like if you were sending it, you know, a massive file to read through because it also changes the way that it reads inputs and obviously outputs as well. You could save up to 60%. So over time, this is saving you money.

In this case, 31% of tokens. And obviously, that'll vary depending on different use cases. Remember, Kayman only shrinks what the agent says, not what the agent does. So, it's still going to operate the exact same. It's not going to dumb down coding, development, scraping, etc., but it's going to make the conversing side of things a lot cheaper and quicker. So, you might be wondering, what's the difference between I have ADHD and caveman?

Well, they're both similar in the sense that they can cut down the length of a response, but I have ADHD primarily focuses on the structure. So, how's it presenting information in what order? With what steps, whereas Caveman purely attacks the volume issue, so you could use both at once. Both do overlap in the sense that they'll both kill fluff, but they attack different problems. If I had to pick one, I'd probably pick ADHD, but if I was trying to save cost, I'd use Caveman.

The best combo that a lot of people have found is using Caveman on light mode. It'll make basic cuts and then use I have ADHD to tackle the structure of your outputs. So, at this point in the video, you now have persistent memory from Claude across all of your sessions. You are now structuring outputs in the correct way that make it easy for you to visualize the next steps. You now have text that actually sounds human and isn't sounding like it's being generated by an AI and you're now saving a bunch of money through caveman mode.

But the final one is much bigger. This addresses a lot of the pitfalls that Andre Kapathy had noticed when he was pushing LLMs to the max when coding. And this is very cool cuz it's just a single claw.mmd file and it improves claude code behavior across the board. So the problem, and this is directly from Andre, is that the models make wrong assumptions on your behalf and just run along with them without checking. They don't manage their confusion.

They don't seek clarifications. They don't surface inconsistencies. They're basically yesmen. They don't present trade-offs. They just they don't push back when they actually should. This is a major problem with AIS at the moment. They really like to over complicate code and APIs, bloat abstractions. They don't clean up dead code. They implement a bloated construction over a thousand lines when 100 would do. And obviously, if you're not an actual software engineer, you probably don't even pick up on this stuff, but obviously Andre is.

They still sometimes even change or remove comments in code they don't sufficiently understand as side effects, even if orthogonal to the task. So, the solution is four principles in a single file that directly addresses these issues. Principle one, think before coding. Wrong assumptions, hidden confusion, missing trade-offs. Principle two, simplicity first, over complication, bloated abstractions. Principle three, surgical changes, orthogonal edits, touching code you shouldn't.

Principle four, goaldriven execution, leverage through tests first, verifiable success criteria. Now, all of the others you're going to use at certain times. Obviously, the claude mem you're probably using all the time for memory, but I have ADHD you'll toggle on and off. KBAM mode you'll toggle on and off. Humanizer you'll only use it when you're writing text. But this one, this is something that you can use all the time whenever you're coding because these principles remain the same.

And this is also the most highly rated GitHub of this entire video. 214.5,000 stars. So to summarize, it's essentially an MD file, so a note that you leave with Claude that says four things. Don't assume, don't overbuild, don't touch what you didn't ask it to touch, and tell me if you disagree. Very basic things that no matter what you're doing, you will want out of an LLM. And this fixes some of the biggest issues I have with Claude, which is the fact it's a yes man, the fact it overbuilds, the fact it makes assumptions, the fact it does stuff I don't ask for.

And there's no code, there's no install. It's just a file that you drop it to your context file. You can put it into Claude under a custom project modification that you could, you know, use even in the normal chats or you can just have it in your folders and use it as an MD file whenever you're working within a specific folder on Claude Code. And the reason why it's got so many stars on GitHub is because Claude is annoying for a lot of people, clearly, even myself.

And this fixes a lot of those issues. So the installation on this one is quite simple. Remember in my school community, I'll have all of the repos in a single MD file. So you can just drag and drop into claude. Some are more complicated to install than others. They'll just all install at once. Free community. Come and join. It's an amazing hub to learn about AI, build a business with AI. We're all, you know, super AI crazy there.

So you can download it from GitHub. And then all you really need to do is drop the claw.md file into the folder that you're working in. So whenever you're working in a folder, it'll have access to this file. You can see the file just states the rules that Claude should follow that we went through. And now whenever I talk to Claude, it will follow these instructions. So even if you're just in the Claude chat or you're in co-work and you just want to use projects, you can actually copy and paste this into instructions.

So here's the MD file. Let's just copy it all. Let's put it into instructions and it's going to save it in instructions. This is like a lazy way to do it. Obviously, if you're using a folder, you just put it in the folder that you're using. You could also just add it to the context folder, but it's going to be less effective than adding it into instructions. So it's really that simple. But clearly this is something that works better if you have it in a folder and if you're just using that folder to work in, which is what I recommend anyway because that's the only way to have real control over your own memory.

Even claude mem that I showed you today is going to guess what it should remember and doesn't give you full sovereignty, which is good for convenience. You want to have it running, but then you probably want to go over and above and save the stuff that you want to save just so you own your own memory. So, those are the five GitHub repos, the five hacks that make Claude not only more usable in 2026, but genuinely turned it into, I think, the best LLM.

I mean, I really like Codeex as well, but the way I have my Claude set up with all my skills and all my repos, it works really well for me. It's really good at writing for me and I still get better outputs than Codeex, unless I'm doing like deep agentic coding work, then I do find Codex is better. But there's a new Chord model launching very soon as well. So, you know, there's always going to be this race between what's better.

One week Codeex might technically be better and then in a couple weeks Chord might be better. But if you are using Claude, using these five repos, I think is going to help you out a ton and actually make Claude smart again. I'll see you in the next video. Subscribe if you enjoyed this video. In the school community, I'll leave the one prompt setup guide for all of these repos, so you're good to go. See you in the next one.

Peace out.