Agent Skills: Practical Way to Cut LLM Context Bloat (Full Transcript)

[00:00:00] Speaker 1: Agent skills are one of the highest leveraged ways to use large language models. It provides scoped task-specific context in a repeatable way, which is exactly what LLMs need to make a reliable output. And the best part? Companies are writing these skills for you. So all you have to do is bring them into your projects and use the workflows. In this guide, I'm going to show you how to leverage these skills to write features that actually work. So what are skills? A skill is a folder with markdown files about how to do a specific task. The main skill must be named skill.md. And the contents of this can be whatever you want, as long as you have the name and description fields in the front matter. Now there are some other optional front matter fields like license, compatibility, metadata, and allowed tools. And you can find them here. But the minimum, you need the name and the description. Now each skill is recommended to be less than 500 lines. In this one, we have around 200. That's because you don't want to load too much context in. But within this entire skill folder, you can have more information in other folders and then reference it within your main skill document. The main skill one should be very focused on the main things you need to do for that skill. The optional nested folders you can have within your skill folder is a scripts folder for executable code, a references folder for any additional documentation that you might need, and an assets folder for some static resources. So what makes these skills so special? Well, one of the biggest problems that comes with AI is called context bloat. Now context bloat is when you provide an AI too much information that it struggles to accomplish an actual task. Like a good analogy for this is a junior developer. Like imagine you just give that developer a specific documentation about whatever they need to do at this moment, and nothing else that's irrelevant, they will probably have a very high likelihood to get a successful result on trying to implement this thing. Now imagine asking that same junior developer to read all the documentation for the whole company. Then after reading it, they need to implement a very specific feature that was a small part of that documentation. The success rate would be much lower. Now LLMs behave the same way. There was earlier approaches such as MCP, they attempted to solve this context problem just by loading a whole bunch of information to your LLM. Now this worked better than having no information, but it still struggled to accomplish tasks consistently because you're just getting too much information. It was too much context bloat. But skills are different. By default, the only thing that gets loaded into the context is the name and description of all the skills that it has access to. Now this takes a very little context, but it gives LLM information about what type of skills are available. Then as your LLM is working through whatever task it's working through, it will decide that one skill is finally necessary. When it decides that a skill can be useful, it will bring that skill MD file into the context window. So now we will have access to all the information within this skill.md. And then as it's going through the skill and trying to understand the task at hand, it will pull in any references that it needs, but only the ones that it actually needs. So this way, your context stays as small as it needs to be for that specific task. Okay, how do you use these skills? Now again, to use it, they're just markdown files. So you just need to bring those markdown files into your project. They live in a subfolder that is usually hidden. So depending on which LLM you're using, so Claude, Codex, or Cursor, they will have the skills within there. For example, the Cursor lives in Cursor slash skills. Codex is Codex slash skills. OpenCode is OpenCode slash skills. Again, these can be copy pasted directly since they're just files, but you can also use skills.sh. Skills.sh is an agent skills directory built by Vercel. It allows you to discover skills based on their popularity. And they also have a CLI tool that they developed to add skills to either your project or your global user. To do this, you just need to run this command npx skills add owner slash repo. Let's do this for the 11 lab skill. So you do npx skills add 11 labs slash skills. So it says it finds there's six skills for 11 labs. We can select all of them. We select which agent we want to install these for. I have Claude, Codex, Cursor, and OpenCode selected. I'll just use those. Do we want to install it to the project or to the user? So you can use it for every project. I'll do it just for this project. And then they have this thing called SimLink. So what SimLink does, we'll just run it and then proceed with installation. Yes. So SimLink makes sure that there's a source of truth for your skills. SimLink puts your skills in this dot agents folder. And in here you can see the actual skill. And I don't know if you noticed the first time, but when we went to Codex folder, it has this little, little icon over here. This means that this is a link. So this agents folder is your source of truth. And all these subsequent folders for different tools are linking to that folder. So this npx skills command is kind of a glorified copy paste, but it does make it a little faster and easier to manage skills. And once you have these skills installed, whether you're using skills at SSH or just copy pasting, using them is very simple. Whichever editor you use should automatically use the skills whenever it is helpful. In theory, once you have the skill in your project, you're done and you shouldn't have to think about it. But at the current moment of writing this, it doesn't always utilize skills at the right moment. So if you know that you need to be using a skill at the given moment, tell the LLM. For example, we just installed the Lemon Lab skills. Now if I want to use that skill, I would just say use the speech to text skill to implement a simple example in this project. Use this as the audio. So now I'll send that message. It said it learned the skill speech to text. Now it's implementing it. So then if we want to use the skill, we just ask the LLM to use it. We say use the speech to text skill to implement a simple example in this project. Use this as the audio. And we'll notice that it found the skill.md file in here. It reviewed it. And now it's thinking about how to actually implement this. We can run the server. And it was able to one shot this demo below. So here's the audio. With a soft and whispery. And we can transcribe it. And this is the correct transcription. There we go. Of course, this UI needs a lot of improvement, but the core 11 labs logic is working and sound all with that single sentence. So I highly recommend leveraging agent skills within your own project. It gives your LM the appropriate context for a specific task curated by the people who made the tools. If you're using 11 labs within your project, install our skills using NPX skills at 11 lab slash skills. Or you can go to the GitHub of 11 lab slash skills. Thanks for watching. Make sure to like and subscribe and I'll see you in the next tutorial. Transcribed by https://otter.ai