Google Antigravity has introduced Agent Skills

Summary: Google Antigravity has introduced Agent Skills—an open standard for teaching AI agents specialized workflows. This tutorial breaks down how to create, deploy, and master these reusable knowledge packages to automate complex engineering tasks.

The promise of AI Agents isn't just "chatting" with code; it's about delegation. But how do you teach an agent your specific team conventions, deployment checklists, or testing protocols?

Enter Agent Skills.

What is an Agent Skill?

Think of a Skill as a "Standard Operating Procedure" (SOP) on steroids. It is a folder containing a SKILL.md file that gives the agent:

1. Context: When to use this skill.
2. Instructions: Step-by-step verified procedures.
3. Resources: Scripts, templates, or examples.

Instead of writing a 50-line prompt every time you need a PR review, you build a Review Skill, and the agent automatically "activates" it when it sees a Pull Request.

1. Where Skills Live

Skills can be scoped to your specific project or available everywhere.

Tip: Start with Workspace Skills. They check into git, ensuring every developer on your team (human or AI) follows the same process.

2. Anatomy of a Skill

Creating a skill is as simple as creating a folder and a markdown file.

Step 1: Create the Structure

To teach your agent how to "Deploy to Staging", create this path:

.agent/skills/
└─── deploy-staging/
    └─── SKILL.md

Step 2: Write the SKILL.md

Every skill must start with YAML frontmatter. This is how the agent discovers it.

---
name: deploy-staging
description: Deploys the current branch to the staging environment. Use when the user asks to "deploy" or "test on staging".
---

# Deploy to Staging

Follow these steps to ensure a safe deployment:

## Prerequisities
1. Ensure the git status is clean.
2. Run `npm run test` to verify no regressions.

## Deployment Steps
1. Run `./scripts/deploy.sh staging`.
2. Wait for the health check to return 200 OK.
3. Notify the user with the staging URL.

3. Real-World Use Case: The "Code Reviewer"

Let's build a robust Code Review skill that forces the agent to check for security and performance, not just syntax.

File: .agent/skills/code-review/SKILL.md

---
name: code-review
description: Reviews code changes for bugs, style, and security issues. Use when reviewing PRs or diffs.
---

# Code Review Protocol

When analyzing code changes, strictly follow this checklist hierarchy:

## 1. Security (Critical)
- **Injection**: Are inputs sanitized?
- **Secrets**: Are api keys hardcoded? (Flag immediately!)

## 2. Performance
- **Loops**: Are there O(n^2) operations on large datasets?
- **Database**: Are we making queries inside loops?

## 3. Style & Maintenance
- **Naming**: Do variable names follow the project's camelCase convention?
- **Comments**: Are complex blocks explained?

## How to Format Feedback
- Use **Bold** for the issue category (e.g., **Security**).
- Provide a `suggested_fix` code block for every issue found.

Now, whenever you ask the agent "Review this file," it will automatically load this protocol and give you a structured, senior-level review.

Best Practices for Mastery

1. Keep it Focused: Don't make a "Do Everything" skill. Make a "Database Migration" skill and a separate "UI Component" skill.
2. Use Scripts as Black Boxes: If a task is complex (e.g., database seeding), write a Python/Bash script and tell the agent to run it, rather than putting the logic in the markdown.
   • Bad: "To seed, first insert users, then posts..."
   • Good: "Run ./scripts/seed_db.sh."
3. Progressive Disclosure: The agent only reads the full method if the description matches the user's need. Write clear, action-oriented descriptions.

Conclusion

Agent Skills transform Google Antigravity from a generic coding assistant into a specialized team member. By defining your workflows as code, you are effectively "programming" the AI's behavior to match your engineering culture.

Start simple. Build a "Project Onboarding" skill today, and let the agent teach your next human hire.