스크립트
00:00:00These are the five open source tools that I wish I knew about when I first started using Claude
00:00:05Code. Because as good as Claude Code is out of the box, it still has some weak spots. Namely,
00:00:12video, front-end design, memory, research, and token output. These are five areas that Claude
00:00:18Code has naturally sort of weakened, and we can improve it drastically by bringing in these outside
00:00:23tools. So in this video, I'm going to show you all five, how you use them, why you should care,
00:00:27and the best part is they're all free. Now, the first tool on our list is Claude Video from Brad
00:00:33Automates. This is at a little over 5,000 stars, so it's a bit smaller, but this is one that's
00:00:38actually trending pretty hard lately. And this is all about giving Claude the ability to ingest video.
00:00:43I don't care about generating AI video. I want Claude to be able to watch videos that I give it,
00:00:48because this is a functionality it does not have out of the box. In fact, the only major AI video
00:00:53model that sort of does is Gemini. And if you're someone who deals with video, you know what a big
00:00:57deal this is, because normally we're stuck only looking at transcripts. And transcripts are great,
00:01:02but sometimes we need the context of what's literally happening on the screen. The transcript
00:01:07isn't enough. But this skill is the best of both worlds, because we not only get the transcript,
00:01:11it's able to intelligently pull screenshots or frames out of the video itself when it deems
00:01:16appropriate. And better yet, if a transcript doesn't exist, it will actually route it through
00:01:20Grok's Whisper model, which is totally free, to generate the transcript. So if you're watching
00:01:24some sort of video that's like a loom or something, it doesn't come with the transcript or whatever it
00:01:28is, we're still okay. Now, the obvious question here is, okay, well, like, how does this actually
00:01:33work? How are we dealing with video? Because we can't just magically have Claude watch videos.
00:01:37Every video is essentially a frame. So am I throwing it 24 screenshots times every second of the video?
00:01:43The answer is no, that would get crazy expensive. Instead, this skill uses a pretty elegant approach
00:01:48where it changes how many frames it's going to grab from the video based on what mode you put it in.
00:01:53So we have four different modes from transcript to token burner transcript. We're not grabbing any frames.
00:01:58We're just taking the captions. For efficient, it's just taking the keyframes. So those are already
00:02:03dictated by the video itself. And we take up to 50, depending on the length of the video.
00:02:07We have balance, which is where a lot of people sit and we'll take up to 100 frames from a video. And
00:02:13it's going to be based on the scene changes, which is also taking a look at the transcript to see if
00:02:18certain words are being used. So balance probably makes the most sense. But we also have token
00:02:23burner, which is essentially the same as balance, but we have no frame cap. You know, you could take
00:02:29a thousand frames. The problem with this obviously is time it takes to do this and the amount of money
00:02:34we're going to be spending. Now, Brad, the guy who created the skill also has his own YouTube channel
00:02:38where he goes into much more technical detail than I am. So if you really want to get, you know,
00:02:43into the down and dirty of how this is operating, definitely take a look at his stuff. As for the
00:02:48installation, it's really simple. You're able to install it into the marketplace, or you can just
00:02:51give Claude Code the URL to this skill. I'll put that in the comment. But big picture, what is his bias?
00:02:56It gives us an entirely new capability that Claude Code normally does not have without us having to do
00:03:01some janky routing through Gemini and essentially pay Gemini API on top of Claude. This keeps it all
00:03:07in-house and is a great addition to your Claude Code stack. Now, before we jump into the next tool,
00:03:12a quick word from today's sponsor, me. So I just released my Claude Code Masterclass and it is
00:03:17the number one way to go from zero to AI dev, especially if you don't come from a technical
00:03:22background. I assume you have no knowledge coming in. We focus on real use cases, and it also includes a
00:03:28codex masterclass. So if you're someone who wants to get a little bit more serious about AI,
00:03:33make sure to check it out. There will be a link to it in the pinned comment.
00:03:37Now, tool number two is all about research because out of the box, the simple web search
00:03:41Claude Code gives us is fine, but it's pretty surface level. And there's really nothing when it comes to a
00:03:46middle ground because the opposite end of the spectrum is let's do dynamic workflows. Let's do deep
00:03:50research. Let's spin up 105 sub-agents and burn up 10 million tokens. I don't want to do that.
00:03:55You probably don't either. So in comes notebooklm-pi. For all intents and purposes, this tool gives us
00:04:02notebooklm inside of Claude Code. I can call on notebooklm through the terminal. Everything I can
00:04:07do in notebooklm from the web version and more can be done through Claude Code because of this skill.
00:04:15It's not just a skill, it's also a CLI. And so it's essentially like an unofficial API
00:04:19into notebooklm. Now, the cool thing about this isn't just like, oh, cool,
00:04:22we get notebooklm functionality, but when you think about it, you're kind of getting free
00:04:26LLM calls doing this. Now it's Gemini. It's not as powerful as something like Opus and certainly
00:04:33Fable, but you can offload some research and some synthesis onto the Google servers for free when you
00:04:38use notebooklm, whether that's just asking questions about videos or whatever, on top of the fact that we
00:04:43can just create, you know, whatever we want using notebooklm, whether that's, you know, slide decks,
00:04:50whether that's infographics, whether that's podcast, et cetera, et cetera, et cetera. And like I alluded
00:04:54to before, we get stuff that goes beyond the web UI itself. And we have a full list right here inside
00:05:00of the readme. In terms of the installation, it has a pretty thorough guide, but I'm going to be honest,
00:05:04all you need to do, copy the URL, throw it in a Claude Code. It's going to do the rest. It's going to
00:05:09require some things like Playwright, which you've never used before, is simply a browser automation
00:05:13that's going to be completely invisible to you when it's running. And lastly, if you really can't think of any
00:05:17use cases of notebooklm, there's a whole list of them right here. For me, the biggest one is simply
00:05:23looking at YouTube videos. And this kind of goes hand in hand with what we were talking about before
00:05:26with being able to watch the videos. The notebooklm is going to be just transcript only, but because
00:05:31it's under the Google umbrella, like it's, it's a very seamless process of supplying it with YouTube
00:05:36URLs, tons of them on a particular topic, and then being able to synthesize all that information
00:05:41at once. Now, tool number three is all about memory. And I'll be throwing in an additional tool here
00:05:46as well. And when we talk about memory, what we're really talking about is how can I have Claude Code
00:05:51quickly and effectively answer questions about very large code bases or very large corpuses of documents?
00:05:59I want to be able to give Claude Code a map that it can very easily traverse to find answers for me about
00:06:05a bunch of different questions that are related to my documents, my work, my code. Well, that is exactly
00:06:10what Graphify does. It essentially creates a knowledge graph of whatever code base you give it. And you see
00:06:15that right here, it breaks out all the parts, it turns them into nodes, it clusters them according to what
00:06:20they're actually about. That way, again, we're handing Claude Code a map. So when we ask questions about
00:06:25things about this code base, there's a very clear path forward from your question to the answer. The thing
00:06:31you need to know, though, is Graphify is not a rag system. There's no vector index, there's no embedding,
00:06:36this is not light rag, this is somewhere in between Obsidian and a true rag system. But we can kind of
00:06:43get like a light version of graph rag if that kind of makes sense. It's not as complicated as traditional
00:06:49rag, yet we're able to get a lot of the same benefits in terms of the memory. Now, the other cool
00:06:53thing about Obsidian versus something like Graphify is it can handle a number of different files,
00:06:58like we're not talking just markdown, we can handle stuff like PDS, we can do images, we can do video
00:07:03and audio on and on and on. So it's very, very flexible. But speaking of Obsidian and knowledge
00:07:07graphs and this sort of thing, let's kind of talk about that bonus tool I alluded to earlier. And that
00:07:11is the Obsidian skills repo. I don't see enough people talking about this. This is actually created
00:07:15by the CEO of Obsidian. It's very simple. It's just a handful of skills. But if you're someone who uses
00:07:21Obsidian with Claude Code, this is an easy way to supercharge it. You're essentially teaching Claude Code the best
00:07:27practices by the people who actually created Obsidian. So don't sleep on this repo, even though it's super
00:07:32simple. That'll be linked below as well. Now tool number four is all about front end design. And that
00:07:36is impeccable. This is quickly becoming my favorite front end design skill. And tons of people are
00:07:41noticing it. It's not just that it has a ton of stars, it's actually like officially part of GitHub's
00:07:45AI package itself. And what we're looking at right here is impeccable's website. And it's here,
00:07:49I'll explain how this tool actually works. So impeccable is one skill, but it has 23 different commands.
00:07:56And you can see all those commands over here on the left that I'm going through things like craft,
00:08:00shape, critique, layout, colorize. And they're essentially having the skill do certain things
00:08:04with your Claude Code setup. So colorize, for example, if I do impeccable colorize, what's going
00:08:09to happen? It's going to add strategic color to monochrome interfaces. What's nice here on the
00:08:13website is I can see a before versus an after. And so you can see, all right, here's what it would
00:08:20normally look like with Claude Code and the standard Claude Code front end design skill
00:08:24versus impeccable. And you can see there's a bit more going on here. It looks a bit nicer. Same
00:08:28thing for, you know, boulder, right? Claude Code, impeccable boulder. And there's 23 different commands
00:08:36here, which are obviously like kind of difficult to explain. And it's much easier just to see them in
00:08:41action. So highly suggest you do that. The other really cool thing with impeccable is the live mode.
00:08:46And this definitely gives you shades of Claude design. The idea is that if I run impeccable
00:08:51live, what's going to happen is it's actually going to bring up my webpage on the local host
00:08:55on my browser. So instead of trying to edit everything through the terminal via code,
00:09:01I will now have the page up on my browser. I can click on different components. I can see what
00:09:05it looks like with and without impeccable. And it becomes a visual design tool, which is way better
00:09:11when we're talking about front design versus like, hey, impeccable, make that look nicer.
00:09:15Okay, try again, make it more premium, right? So you can actually see it before you commit it. I think
00:09:21this is a huge step above the anthropic front end design skill and also a huge step above things
00:09:26like UI, UX, Pro Max. Now, last but not least is Ponytail. And this is all about token consumption,
00:09:33tokens, tokens, tokens. You hear about this all the time and how expensive they are, especially with
00:09:37Fable. So it only makes sense. We look outside of Claude Code to see, are there any skills or frameworks
00:09:43that can reduce the amount of tokens we are spending while still maintaining the same level of effectiveness?
00:09:49You know, it does us no good if we reduce our token count, but it gets worse. Well, Ponytail
00:09:54claims to be able to do this. In fact, it claims that it makes Claude Code 20% cheaper, 27% faster,
00:10:02while still giving the same results, which is kind of wild. Now, the way Ponytail essentially works
00:10:08is it saying, hey, we're going to give Claude Code the series of, you know, gates it needs to pass,
00:10:14where essentially we ask it, hey, do you actually need to build it? Does that feature you're trying
00:10:18to create already exist? Is it a library? Et cetera, et cetera, et cetera, before finally saying,
00:10:23okay, you want to build this? Great. Thumbs up. Just use the least amount of code.
00:10:27That's kind of how it works in a nutshell. It gets a little more complicated than that.
00:10:31But I wanted to have you take a look here at the benchmarks because this is what we care about.
00:10:36In the gray, what do we have? We have the baseline. And then in the green, we have Ponytail. And you
00:10:41can see way less lines of code, way less tokens, way cheaper, and way less time. Now, what is the catch?
00:10:49Well, the catch is they did these benchmarks with Haiku. You at this point are using Opus or you're
00:10:53using Fable. So does this hold up? Well, I actually did tests. I did a whole video on Ponytail with Opus
00:11:00and it actually was even cheaper and quicker than what we see with Haiku. So the benefits were greater
00:11:07with Opus. I then tried it again with Fable in the same thing. So across the board, when I ran these
00:11:13same benchmarks, and anyone can, if you go on this repo, they have all the benchmarks listed here so you
00:11:17can test this yourself. Ponytail reduced it and it was the same output. Now, benchmarks versus real
00:11:23life. Is it the same? Who's to say? It probably depends on your particular use case and how
00:11:28complicated it is. But any chance we can make Cloud Code faster and cheaper and at the same level of
00:11:35effectiveness, I think we should try it out. Worst case scenario, you do a couple of runs, you don't
00:11:38like it, you get rid of it. But I think this is worth your time. There's other ones in the same
00:11:42vein, like Caveman, that I also think you should take a look at. So those are the five open source
00:11:46tools that I wish I knew about when I first started with Cloud Code. And if you're brand new, I hope I was able to
00:11:52at least point you in the right direction in a few of these areas. As always, let me know what you
00:11:57thought in the comments. Make sure to check out Chase AI Plus if you want to get your hands on the
00:12:00masterclass. And besides that, I'll see you around.