Transcription
Agent skills, cloth skills, sub agents, MCPs. What is going on in this world? Let me explain these concepts to you in a very simple way, and I'll show you a few examples to guide you how to actually use this in your day-to-day life.
For those of you who don't understand what is an agent skill, think of something like you have a personal assistant, and perhaps their job is to take down the notes after every meeting. And you want your notes to be in the exact same format for every single meeting you have with your business partners. What this personal assistant needs to do is that they would need to understand what exactly are the things that you are interested in reading in this report, so that they can structure it in a way that will follow your own style. That is why skills come in handy when we are trying to use claw code for uh doing some repeated tasks.
And before I show an example, Enthropic also launched this concept called MCP a while ago. So the key difference between these two is actually very simple. For MCP, the way you do things is that you're asking for information. You're asking it to have to get access to some documentations from a different platform. Again, using the same personal assistant example, you might want this personal assistant to go to the storage room to find out some old documents for you and then fetch it and bring it to your table. So, that's this is what MCP does. I made an MCP video a while ago which was explaining how we can use MCP to fetch information from Notion notes. It's basically something that you need to read information. And it's like an API for software, but it's just for AI agents.
Instead, what skill does is that it can take it to the next level and be like, okay, after you fetch this information, for instance, it could be local, it could be fetched by MCP, it could be fetched by an API, how can you guide your AI to process that information in a way that you want, just like the example we explained earlier. So here, what really matters is you need to define a file called skill of how you want to process information.
Some of you might also be wondering, what is a sub agent? It's basically like who's doing the task and what they should be doing, right? For people who are familiar with a deep research agent, remember like when you ask a question, they would always clarify with you first. That's because there's a supervisor agent there trying to, you know, understand what your intent is, and then they're going to assign the task to all their little sub agents. We call it researchers. Do the individual research. So here, what's going on is that claw code is going to break down the task that we're going to give to them. For this video, we're not going to focus on sub agents. We're going to focus on how can we teach claw code right here to be my best personal assistant that understands my styles.
Okay. So let's jump into an example right now. So in this git repo launch agent skills, I prepared a sample WhatsApp chat script. This WhatsApp chat script has 74 rows and it's basically a sales process between a client and the business talking about a B2B sales. Okay. So, as a manager, I would probably want to know what is going on in that conversation and what are some of the to-dos that we should be uh looking at, right? And imagine if your sales team have like a 100 conversations like this. Yeah, you you would have no time to understand this, but you want a structured way to document all of these u to-dos and u important information. And that is where claw skills or agent skills will come in handy.
All right, let me show you real quick. So, what I would do in claw code is that I'll just be typing, "Help me summarize the sample WhatsApp chat text. Let's see what's going on." So, it says it's pondering, it's cooking. You see, it read the skill file. If I click on it, it's called skill.md. And it was reading this specific skill called WhatsApp summarizer. Let's just take a quick look at this skill document before we go any further.
So essentially the structure of this document is actually well defined. The first section is called a metadata section, and it's always loaded right. And it the claw clot code will basically scan these to find the relevant skills, right? What you need is a name, a description, and triggers. All right. The name here is that it's a chat summarizer. Description is that it's going to generate concise summaries about um the chats from uh WhatsApp. And then the trigger is that if the user is asking for summarizing the chat or something.
The second part is called the body, where after the user query matches with the trigger keyword, it will start loading this entire thing down below, right? I know I defined a really long document, but it could be as short as, you know, how many times people use the word um vibes, right? That that is another skill uh body instructions that you could do. But here, what we did is that we explained the purpose, which is transform the parse state chat data into actionable summaries. And then there are some prerequisites, and there are clear instructions for how you will be able to summarize. I was also clearly defining, you know, what is the summary structure that you should be following uh if cl code is analyzing my chat. So the structure here that I did is that you should always include like topics, decisions, open questions, deadlines. Which, if you go back to clock code, you can see that, you know, uh topics, uh decisions, you know, action items.
I would say congratulations that you have already seen a really simple example of using agent skills. But what can we do even more? There's this third section which is called resources. What resources does is that it's also able to reference some external files. Like, for example, you might need some code uh to help you finish some task. Like um, what if I would like to uh export this kind of thing into an HTML, right, into a website? Why don't we just try it right now? So for WhatsApp, we also had another skill called action extraction. So what I could do is that, uh also extract the to-dos for WhatsApp chat and export import it to HTML.
All right, while this is running, you can see we have a little folder called scripts, in which there's a Python file called export to HTML.py. Clocko actually read the skill of extracting uh WhatsApp actions, and export to HTML is one of the resources that this agent skill will have access to. Build an HTML file like a website uh in the format that we defined down below. Let's see the result when it's done. Cool. So this is done, and the complete analysis has been exported to sample WhatsApp chat analysis report. I'm just going to say, "Open it for me, please." Yes, wrong command. Boom.
Okay, look at this. So this is a freshly generated report of that WhatsApp chat conversation, completely using claw skills and agent skills. All right, remember this is our exact summaries, which is about a B2B sales conversation between Marcus and Jennifer on WhatsApp. And there's some key decisions here too. The topics that were discussed, the pain points, um deal structure, and everything, uh, and and the high priority to-dos that uh we talked about in the extractions. All right. And eventually, we are talking about the timeline because in the skill, we asked it to summarize the deadlines.
Okay. I think by watching until here, most of you should have got a pretty good understanding of what skills mean to agents to clock code. I would encourage everybody, first of all, to clone this GitHub uh to try with your own hands on claw code. And there are clear instructions in this git repo for readme. And also, I highly recommend to go take a look and read the claude API documentations on agent skills. It has everything that I talked about over here. And I was just thinking that making this video could make it easier for you to understand the concept and see a real working example, but reading this document was very helpful for me personally. Um, it was also talking about the metadata, the instructions, and the resources code. So, I didn't make things up.
Okay. In the rest of the video, I want to show you something slightly more technical because I think a lot of my audience are u probably developers or trying to learn how to build products. And many of my old audience also were watching me showing you how to build things from scratch. So what I think would be really impactful for developer community with agent skills is that every single time when I try to start a new project, I always need to manually configure a bunch of things, and they would be exactly the same process every single time. For example, I had this uh repo called a launch rag. Let me just open it for you. Uh, which is helping people to launch an AI agent rag backend using fast API immediately. And if you scroll down a little bit, you can see that, oh, there are a bunch of steps, you know, got to clone, and got to set up the API keys for superbase, got to configure the environment, everything. So, I was thinking, hm, why why can't we use agent skills to get it done? Must be very fun and interesting. And uh, I built something just for that. And let's dive in and take a look.
So, so the goal is basically we want to launch a agent backend with retrieval augmented generation. Uh, I had a previous video about this. Feel free to check it out about um in the exact like logic behind the entire project, but that is not the focus for this video. This video is about how can we dive in to this uh agent skills workflow to set it up without me even touching it. Okay. So what I prepared is that if you remember, we kept the WhatsApp skills in this skills folder. All right, we have a few WhatsApp skills. Let me just hide them. Uh, we also have thing called rack database, rack setup, rack tools. If I click into rack setup, you can see that uh we also have the metadata on the very top. It's basically setting up a fast API project. The description is that you got to clone this project and set it up. Um, and then the trigger is basically fast API API setup blah blah blah. What's more important is the instructions. How do we set up uh an instruction so that it will follow it step by step?
Why don't we just try it right now? So, I'm starting a new claw code and I'm just going to ask this question, "Set up a fast API backend for my AI agent project using launch rag." Okay, hit send. Okay, so it's reading the skill, right? It's reading the rack setup skill. Okay, so now it knows it's asking me if I should do launch rag or launch agentic rag because I clearly set out these instructions in this onboarding flow to ask the user which version I want to set up. So I would just say, "Okay, let's do one because it's simpler and I just want to make it start working." Uh, "Perfect. I'll set up launch rack." Um, and then "Current directory is this or different location?" So, "Current directory is fine." "What would you like to name your project folder?" I'll just say, "rag fast API YouTube demo." You can see that I have this old project here called old fast a project. So you should be setting up something new for me. You see, if you're a developer, you will understand what I'm trying to say. Sometimes there are things that we just want to um be automated without me, you know, thinking too much about it because it's a repeated process. Things like superbase usually would do that in front end of superbase. But in this case, I'm just trying to tell it to set up for me completely using the terminal, using command lines.
It was asking me, "Would you like to help me set up the database setup? Set up the rack database for me too?" And then it read a different skill called rag database skill. "Yes. CLI only." Yeah, it's asking me for the project names, right? As we expected. So, I'm just going to say, "rag claude agent skills demo live." [snorts] So, it looks like they have already generated this new project for me. We do have a file to initialize this table called create table rack chunks. Let me just refresh this. Good. So it looks like the table has been generated just like before. We have all the column names: ID, chunk ID, source text, embedding, created at. So this is good.
Last but not least, let's try the chat. There we go. Look at this. So now is the demo time. Uh, for those of you who are not familiar, I previously made a video about this. Uh, essentially on the right hand side, you have some policies for this e-commerce chat where the policies are supposed to be embedded into vectors so that when a customer is asking questions to this SI system [snorts], then we're going to do a round of vector search for, you know, the embeddings. So the first step we're going to do is called seeding. Right now, it's a fresh project, and if we go to superbase right here, you can see that the table is empty. Okay. So what we got to is that if we just click on save here, after we I click on save, I come back. I refresh. Good. It's working. I didn't touch it. It created this project for me. It embedded these policies for me. And one of the policies is that there's free shipping for over $50. If I ask, "I bought a pair of shoes for $49. Can I ship it for free?" So now it's sort of doing the vector search uh embedded in my question, doing a match to the policy, and it say, "Free sandwich shopping is available for $50 plus. So your purchase $49 is not qualifying." Say, "Anything over $20 is for free for shipping." If I click on save, what it did is that it also updated this database for embedding. And if I ask the same question, "Yes, you can ship your $49 pair dollars pair of shoes for free."
So that was an example of setting up this entire rack backend system and then test it real quick, and at the same time setting up Superbase uh without me touching the front end at all. I think this is amazing. Maybe somebody could create something that will send you notifications on your phone with a pop-up notification. If you click into it, you can just approve uh give it more command, and it's going to continue to run. Man, the future is already here. It's just not evenly distributed. I think this is the best time to build startups.
And the reason why I chose WhatsApp as an example to showcase agent skills is that we're building automatis, which is an AI sales leads agent platform for business who are selling customized goods and services. And uh, we allow people to uh automatically capture leads through WhatsApp chats, uh web chats, emails, and turn into structured leads so that you won't lose business opportunities. Or if you're putting meta ads and then you get like thousands of inbounds, you'll have an agent to take care of it for you. You don't need to do the manual task on CRM. That was the ads. Uh, thanks for listening. I hope my explanations is clear, and let me know if you have any questions. Feel free to join our Discord or just DM me or comment down below. I'll see you next time. Pex.