Intro to Agent Skills

AI progress moves at an incredible rate and standards, best practices, and techniques are constantly evolving.

Sometimes we make big leaps with the models, but other times the tooling can close critical gaps. We've moved away from simple code completion to full agentic driven chats. But we've hit a

wall, context bloat. You can't fit your entire codebase into the context window without confusing the model or wasting tokens. Agent skills are a great way to

tokens. Agent skills are a great way to provide narrow and deep context for how to do something, but allow the agent to discover them only when they're needed.

Because they can be lazily loaded in, you don't have to pay the cost of the context bloat for the entire skill definition if it's not relevant to the task at hand. At its core, a skill is

simply a skill.md file in a named folder inside the agent skills local or global directory. The skill.md file is just a

directory. The skill.md file is just a markdown file with front matter that includes the name and description that the agent will use to see if the skill makes sense to read in for the current

task. You can have global skills that

task. You can have global skills that apply to the frameworks and languages you work on and project specific skills for libraries, SDKs, and frameworks specific to the app or backend that

you're working on currently. A skill can include other resources, scripts, code examples, and other assets that are needed. It makes it easy to colllocate

needed. It makes it easy to colllocate and distribute the skills with your team and manage them on disk. You can find skills in many popular repositories from GitHub repos or even have the agent

generate them for you.

Let's walk through how to build custom skills together. First, we can open

skills together. First, we can open anti-gravity and create a new project in the playground. Create a nested folder

the playground. Create a nested folder called aagent/skills/html3danvas and create a skill.md inside of it.

Inside the skillmd, we create triple dashes to indicate the front matter region and provide a YAML name and description. We then can ask Gemini in

description. We then can ask Gemini in the agent chat to generate us a skill for working with 3GS and V and how to set up the resize listener as well as the animation loop. We should also ask

for a skill for working with the new TSL tiny shader language, which is a recent addition to 3JS and may not be well represented in the training data. After

it creates the skills, we then can start a new chat and ask the agent to build us a 3D solar system on the web. The agent

should come up with a plan and while generating it should read the skills and load them into context. Agents are

getting better with training data and sometimes it'll be able to solve the problem without skills. But by providing the context, you can influence the specific type and style of output that

you're going for and better convey the intent for when you're asking for various tasks. These can also be very

various tasks. These can also be very company and industry specific and you will add more as your project grows. Now

that we have a project, we then can ask the agent to create a skill for building games with 3JS and another skill for adding HTML guey that sits above the canvas. After that is done, we can start

canvas. After that is done, we can start a new chat. We can ask the agent to update our code and create a learning based game for the solar system. Then we

can have a generated quiz and facts for each planet that we click on. As it

generates, you should see it read in the new skills that we created. And when it is done, you can see if the game works.

When we decide to share this code with others, this will load the project into anti-gravity and the skills will be discoverable for them, making distributed knowledge possible, meaning that they can have it colllocated with

the project. There are many places to

the project. There are many places to find skills and many tools that are being created to make it easier. You can

use skills.sh sh to add any number of skills for supported git repos. For

example, skills are an open standard and you can check them out at agentskills.io.

agentskills.io.

You can get started today with anti-gravity and skills by downloading the IDE and adding your favorite ones.

We'll include a link to some good skills to check out in the description. Thanks

for watching and let us know how you're using skills to level up your agentled development. Thanks.

development. Thanks.