🌿AI Agent Skills Complete Tutorial: The Standard Way to Give AI New Abilities

📝 About This Article

I'll be honest with you. The author is me, but the writer is AI. The inspiration, knowledge framework, and research come from me personally, while AI helped me organize these materials into article form.

This is also a practice of the "AI collaboration" spirit that this article explores.

Introduction

Since 2025, the concept of Skills has become popular worldwide.

It was first proposed by Claude Code, but now all major companies have adopted it — or should I say "paid tribute" to it?

In any case, the moat in the AI industry is quite low, and things can be imitated or surpassed at any time.

When using AI Coding tools, there's a common frustration: AI often lacks sufficient background information and doesn't know your team's habits and workflows.

For example, you might ask AI to help deploy to Vercel, but it gives you steps that are completely different from your team's process.

This is where Agent Skills come in handy!

This article will introduce this important concept in 2026 AI development from scratch, helping you understand everything from concepts to installation.

What Are Agent Skills?

Simply put, Agent Skills are "skill books" or "Standard Operating Procedures (SOPs)" for AI Agents.

It's a simple, open format that enables Agents to acquire new skills and expertise.

Skill Components

A Skill is typically a folder that may contain:

Once you install a Skill, the Agent will automatically consult this SOP when encountering related problems, instead of "hallucinating."

Why Do We Need Agent Skills?

You might ask: Isn't AI already powerful enough? Why do we need additional skills?

The answer is: No matter how powerful an AI model is, it lacks your team's unique background information.

Value for Different Roles

What Does Skill.md Look Like?

A standard SKILL.md is essentially a Markdown file containing two key parts:

1. Frontmatter (Metadata) - For the IDE, describing skill name and purpose
2. Body - System Prompt for the AI

Example: SOLID Principles Review Skill

---
name: SOLID Code Reviewer
description: Analyze C# code for SOLID principle violations.
version: 1.0.0
---

# Role
You are a Senior .NET Architect specializing in software design patterns.

# Objective
Review the provided C# code fragments and identify violations of SOLID principles.

# Output Format
Please provide the analysis in a table format:
| Principle | Violation Detected | Suggestion |
| :--- | :--- | :--- |
| Single Responsibility | ... | ... |

# Constraints
- Use Traditional Chinese (Taiwan) for explanations.
- Do not rewrite the code unless asked.

Structure Breakdown

• Frontmatter (--- block): Tells the system what this skill is called and what it does
• Body: When you invoke this skill, this text is dynamically injected into the AI's Context Window

How Does AI Know to Read SKILL.md?

This is a great question about LLM Context Management.

Key point: This is definitely NOT fine-tuning or training!

Models like Gemini, GPT, and Claude have no idea about your SKILL.md when they ship.

Antigravity makes AI "appear to understand" your skills through a hybrid technique of RAG + System Prompt Injection:

Phase 1: Skill Discovery

1. When the IDE starts, the backend scans the .agent/skills/ folder
2. Reads the Metadata from each Markdown file
3. Stuffs the skill list into the Agent's System Prompt

Phase 2: Dynamic Injection

When you ask: "Help me check if this code follows SOLID principles?"

1. Intent Recognition: LLM finds the question matches solid-reviewer
2. Tool Call: LLM requests to view the skill's detailed content
3. Content Injection: IDE pastes the complete SKILL.md content into the Context
4. Final Response: LLM can now respond according to your SOP

Why Not Just Train It In?

What Is Antigravity?

When talking about Agent Skills, we must mention Google Antigravity — this is the new generation Agent-first IDE launched by Google in late 2025 (modified based on VS Code architecture).

Antigravity is not just a code editor; it has a powerful built-in Agent Manager.

The AI here doesn't just autocomplete code like GitHub Copilot — it can "autonomously" plan tasks, execute terminal commands, and even operate browsers.

Core Concept

You're no longer just writing Code — you are the "Commander," and the Agent in Antigravity is your "Executive Officer."

How to Install Agent Skills?

Method 1: Automatic Installation Using CLI (Recommended)

Using Vercel's official skill library as an example, run in Terminal:

npx -y skills add vercel-labs/agent-skills

This command automatically detects your environment and installs Skills to the corresponding path:

Method 2: Manual Installation

If you want full control, you can create manually:

# 1. Create folder
mkdir -p .agent/skills

# 2. Create skill file
# Add .md file under .agent/skills/ (e.g., csharp-expert.md)

# 3. Restart environment
# Press Ctrl+Shift+P, type "Reload Window"

Automatic vs Manual Comparison

Differences Between VS Code and Antigravity

Although both use npx skills add ..., the paths and activation mechanisms differ:

Plugin Marketplace (Advanced)

Beyond Skills, there's a more complete Plugin ecosystem:

What Is a Plugin?

Plugins can be packaged as bundles containing:

• Skills - Let Agents use your team's workflows
• MCP Servers - Auto-load tools after installation
• Custom Agents - With specific roles and instruction sets

Example with Claude Code

# Add marketplace
/plugin marketplace add microsoftdocs/mcp

# Browse plugins
/plugin marketplace browse anthropic-agent-skills

# Install plugin
/plugin install document-skills@anthropic-agent-skills

Professional Advice: Hybrid Strategy

Since you understand architecture, I recommend a hybrid approach:

1. First use npx to grab official Skills as your "base class"
2. Then go to .agent/skills/ and directly modify the Markdown
3. "Override" according to personal preferences (e.g., force use of Traditional Chinese)

This follows the Open-Closed Principle — customize while standing on the shoulders of giants!

Summary

Antigravity is essentially a powerful Prompt Engineer Agent.

It helps you manage these Prompts (Skills), feeding the right Prompt to the LLM at the right time.

This is why path names are so important — the IDE's program logic is hardcoded to scan specific locations. If it can't find them, it can't feed them to the LLM, and the LLM reverts to being a generic model that "doesn't know your rules."

Additional Resources

• Vercel Official Agent Skills Collection
• Claude Code Plugins Documentation

• first edition time: 2026-02-08
  

This is a growing article that will be continuously updated. Subscribe to my blog for the latest information!