00:00Okay, so Anthropic recently released skills for Claude code, and according to some people, it's even bigger than MCP

00:05Well, we'll have to see that

00:11I've been testing it out for the last 3 to 4 days and digging into how it works and where it shines

00:16So in this video, I'll share everything that I've learned so far, and we'll cover the core idea behind Claude skills, how it compares to sub-agents and projects

00:25In fact, I've created an entire chart to understand that

00:27I'll share it with you, how to build your own custom skills, context management using skills, and finally, we'll wrap it up with a hands-on example

00:35So, Let's dive in and figure out what Claude skills are all about

00:39According to Anthropic, you can think of skills as custom onboarding materials that lets you package expertise, making Claude a specialist on what matters most to you

00:48Simply put, skills are like folders

00:50They contain instructions, scripts, and resources that the agent can use to perform specific tasks

00:56For example, if you're using Claude code to create some GIFs for Slack, you probably follow the same steps every time, write instruction, provide examples, repeat the whole process over and over again

01:06Now, what if you could bundle all These instructions, examples inside a GIF file, and upload it to Cloud Code and reuse it whenever you need

01:14So, all you need to do is to go to Cloud settings capabilities, and in here, you have skills available as preview

01:22So, either you can upload a new skill, which we are going to see in detail as we move ahead in this video, or you can use some built-in skills as well

01:30Some of the examples can be creating an algorithmic art, creating uh brand guidelines, and stuff like that

01:36So, similarly, you have skills for building MCP server, and even there's a skill that you can use uh to create new skills

01:39We are going to see

01:41That as well

01:44So, uh, in here we already have a skill called, uh, Slack GIF creator

01:47So let's enable that, and this should give you an idea of, like, what exactly a skill can do for you

01:54So let's try that in chat, and it says that, "Hey Claude, I just added the Slack GIF creator skill

01:59Can you make something amazing with it?" So let's see how it goes

02:05So now Claude is, uh, reading all the documentation for this skill, and it has decided to create a bouncing, rotating star with sparkle particles

02:11So let's see how it goes

02:18Okay, so this is what it has created for us

02:20It's a fairly simple example, but I'll show you how you can use Skills to automate any repeatable workflow that you use every day

02:27Now, a question might arise, why do we even need skills when we already have projects, MCP, and sub-agents? Well, to answer that, it's important to understand a bit more about skills, how they're created, and where exactly they fit in, so let's walk through that first

02:41So currently, you're looking at the overall architecture of how a cloud agent works

02:44An agent can be equipped with different skills, and the skills then can be copied to the agent's virtual machine

02:52Now, a skill is basically a directory that contains a skill.mmd file

02:56Basically, some instruction about how you can accomplish a particular task, along with that, it can have some resources

03:01It can have some scripts that are important for the agent to accomplish that particular task, right? So all of these are going to be copied to the agent's virtual machine, and then the agent can use that

03:12Now, it's important to note, where does MCP fit in here? Well, MCP is still an integral part of your agent's architecture for tool calling and connecting to external services, whereas skill is more like a set of instruction resources and nuances, how you can accomplish a particular task

03:30So they basically complement each other, making the agent even more robust and stronger

03:36So a natural question now is, what is so special about these skills, and how does Claude decide which skill to actually use? To answer that, let's understand the anatomy of a skill

03:46So a skill is nothing but a simple markdown file

03:48It starts with a front matter, right? So it's a brief description about the skill, like what it is supposed to do

03:56I'll tell you the reason, like why all of this division is being done

04:00Then it is followed by a markdown description of what exactly the skill is supposed to do

04:04So this basically is the

04:09actual body or the content of the skill that, uh, tells more about, uh, what you can use this skill for

04:13Along with that, the markdown file can have references to additional resources or documents, since it has full capability to have its own examples, own documentation, own code files that it can execute to accomplish a particular task

04:31So in case, like, uh, if the entire context gets very longer, you can easily, you know, create these, uh, reference files, resources, or code scripts that can be used by the skills, and they can be referenced directly inside the markdown file

04:46Now you must be thinking, why do we need to do All of this? Imagine you have got several skills available, and a new task comes in

04:54So instead of the LM loading every skill in full along with their metadata, which would bloat the entire context, it only loads the front matter description

05:04The front matter description is something that we saw here

05:05So it briefly tells the agent that, okay, this is the particular skill that matches with the task at hand

05:12And once that happens, when we actually trigger or load the skill, then we're going to, you know, load the entire markdown or the entire content of that skill, followed by that as the skill is being executed

05:24Or, as the task is being executed, on the need basis, we are going to load all the resources like text files, scripts, and data

05:31So, although design-wise it's fairly simple, it's really powerful when it comes to managing larger context

05:38Now that you understand what skills are, and before we jump into a hands-on example of uh building our own skills, it's a good time to go over some of the key differences between skills, projects, and sub-agents

05:52So, I've created this table for you, which I'm going to attach below as well, but if I quickly summarize the fundamental difference uh between these three, skills are Reusable procedures that load can automatically invoke when a relevant task is encountered, while projects are workspaces that you choose to keep related files and conversation organized in one place

06:12And lastly, sub-agents are separate instances of code that can handle specific subtasks independently, each focusing on its own piece of work without mixing things up, and each has its own context that is self-managed

06:27And another key difference is how these three are invoked

06:29So skills are automatically invoked by Claude code based on the task, and projects are manually selected by user

06:36So, They go into a project and do whatever the conversation or whatever the task that they are trying to do inside that, and sub-agents are something that are invoked by the main Claude agent to delegate and assign a subtask

06:46So I hope this clarifies, like, why skills exist and where exactly they fit in where we already have projects and sub-agents

06:53Now let's jump into the demo, and it's time to build our own skill and learn how we can use it with either Claude code or Claude desktop

06:59Okay

07:05So now I'm inside my cursor IDE, and we are going to see how you can build your own custom skills and use it with either Claude code or Claude

07:08desktop

07:12So if you are inside Claude code, I mean it would hardly matter how you create the skills

07:15It's just that the interfaces are different

07:17So if you're inside Claude code, uh, we already know that, uh, the dot Claude folder has, uh, agents commands, and, uh, similarly, you can create a folder for skill, and inside that, you know, you can add as many skills as you want

07:30So let's look at this one, which is menu visualization, and, uh, every skill has to have a skill.md file, and, uh, this is what we have, and we already understood, uh, the anatomy of how to create a skill

07:44It should start with a YAML front matter, which is basically a

07:48Very small description of what the skill can do, and this is something that is always loaded into context to understand whether the skill has to be invoked or not, instead of, you know, loading everything inside the context window and blotting it up

07:59So this is how the context is being managed by Claude code, and then it has a markdown structure, or the actual body of the skill, which contains all the instruction about what it can do

08:11So some of you might already know that Manim is basically a library to create three blue brown like mathematical animations

08:20So we have created a skill for that

08:23Bundled all the resources and everything that one needs to, or the agent needs to, create similar visualization

08:29We'll see a hands-on demo of how it works as well

08:31And along with that, uh, if you have some additional resources that you want to bundle together with the skill, then you can provide references to those inside this skill.md file

08:42So in our case, we have, uh, these, uh, scripts examples.py, so which is being referred here in the skill.md, and we have some references to mathematical objects file, so which is here

08:55So these are basically some of the basic mathematical object that are going to be used by our

08:59library or used by our skill, and along with that, we have also provided some standard examples of uh, how one can create or use uh, Manim for creating this visualization, so you can see some of the examples here

09:14There's an example for equation, there's an example for uh, constructing graph and stuff like that

09:19So these are going to be the references that the agent can use whenever a new task is uh, encountered, and it can pick some ideas from here

09:27Now that uh, you have your skill ready, you can directly invoke it by saying that, okay, let's say create a sine wave animation for me using the Manim ID skills

09:37So it's Fairly simple to do that, but I think it would be more fun if we try to use it with Claude desktop, and uh, let's try to do that

09:43So once you have created the skill, you can, you know, uh, just compress it, uh, in form of a GIF file, go to Claude code, and uh, so once you're inside the Claude, you just need to go to settings, capabilities, and inside that, we have, uh, the skills here

10:03So you can directly upload, uh, one of the skills here

10:05So let's grab one of the skills that we created

10:08I already jipped the manim with skill that we have, and I just need to drop it here

10:15So since I've already uploaded this skill, uh, either I can replace it, or Just cancel it

10:20And, uh, since, uh, this skill is already activated here, it's time to, you know, quickly try it out in chat

10:27So let's do that

10:30So it says that, "Hey, Claude, uh, I just added the Manim with skill

10:32Can you make something amazing with it?" So let's see what it does for us

10:35Okay

10:39So it has, uh, decided to create Euler's formula for us

10:42So let's see how it goes

10:46And as you can see, uh, it is trying to install Manim, which is the library that is going to be used

10:50So basically, your skill can encapsulate all the resources, instruction, along with the code scripts and instruction on how to install them

10:58All of that can be Encapsulated or bundled inside your skill

11:02So, which we have done and, uh, the agent is trying to leverage all of the data

11:06All right

11:11So it has finally finished, uh, creating this, uh, visualization for us and, uh, let's open it in QuickTime and, uh, see how it works

11:14All right

11:17So let's see that

11:20So, I mean, this is pretty much, uh, what we see in many of the Three Blue One Brown YouTube videos and, uh, nice

11:30So this is some really nice animation

11:31Now, the thing is, this skill helps you create these many diagrams again and again, like you do not have to repeat yourself with providing all the instruction and everything that, uh, Uh, the agent would need to, you know, accomplish these tasks

11:46So, it's uh, just like you have to sit once, put down all the instructions about any task that you're trying to do, and uh, bundle in all the resources, scripts that are required there, and you just upload it to Claude code or Claude desktop, and uh, you can use the same workflow again and again

12:03So, before we wrap up, uh, I quickly wanted to show you one more thing

12:07So, let's say how you must think about creating your own skill

12:10So, what you can do is, uh, you just need to think about, like, what are the tasks that you are, or what are the workflows that you are using every day, and uh, How you can automate them, right? So every time maybe you're using Claude desktop or Claude code, you must be writing a bunch of instructions and all, right? So let's say if I want to, you know, create a skill for writing YouTube scripts, so Claude desktop has a built-in skill, and this skill is known as a skill creator skill

12:38So basically, it has all the guidelines on how you can create your own custom skills, so let's try to use that in chat, right? So you can start with some bare minimum instruction

12:56It can always keep on asking you questions to refine what you require the skill to be

13:00So now it is trying to clarify

13:03More about what the skills has to be, so what your typical video style is, so whether it's educational product reviews and stuff like that

13:12So let's say it's uh educational explainers

13:14And the next thing is, uh, so include all of these

13:22We'll say include all of the following

13:24And, uh, so now it says that, uh, it has everything that it needs to, you know, create the skill for us, right? So it has, uh, created the skill for us, and we can also ask it to, you know, so we are asking it to, you know, uh, jip this, and, uh, something that can be used similar to how we uploaded the manimage skill to our Claude code, right? And, uh, I can just

13:49Download it from here and upload it to, uh, my existing set of skills inside Flow Desktop and use it over and over again, so it's just a fairly simple example

13:57Like, if you are trying to, you know, uh, build something of your own, then you can provide additional examples and some more, uh, let's say code files, uh, reference docs, and all

14:06So all of that can be done

14:08It only takes, uh, you know, a bit of hard work in the beginning, but once you're done with this, I think this is something that can save you tons of hours because you would never have to repeat yourself

14:16So with that, uh, we have come to the end of this video, and, uh, thank you so Much for sticking around till the end, and uh, as always, I'll see you in the next one, until then, bye-bye.