What Are AI Agent Skills? A Guide for Antigravity

An AI agent skill is a folder with a SKILL.md file that tells a coding agent how to do one kind of task well. Google's Antigravity docs describe it as "a folder containing a SKILL.md file with instructions that the agent can follow when working on specific tasks." The agent reads a short description of every skill up front and loads the full instructions only when a task calls for them.

This guide is for developers using Google Antigravity. It explains what is inside a skill, how the agent loads one, how skills differ from rules and MCP servers, and why you should read a skill before installing it. The Antigravity specifics below come from the official skills documentation at antigravity.google/docs/skills.

What a skill contains

Skills are described as an open standard. A skill holds instructions, best practices and, optionally, scripts and resources. In Antigravity a skill folder looks like this:

my-skill/
  SKILL.md      # required
  scripts/      # optional
  examples/     # optional
  resources/    # optional

Only SKILL.md is required. It starts with frontmatter, followed by the instructions in Markdown:

---
name: my-skill
description: Audits a page for on-page SEO issues and ranks them by impact.
---

# Steps
1. Ask for the URL or pasted HTML.
2. Report findings in priority order.

The Antigravity docs list two frontmatter fields. name is optional. description is required, and the agent uses it to decide whether a skill applies to the task in front of it. A vague description means the skill rarely gets picked; a precise one means it fires at the right moment.

The optional folders hold the heavier material. scripts/ is for code the agent can run, examples/ for reference implementations, and resources/ for reference files such as templates or checklists. Many skills are only a single Markdown file with no code at all. If you want to write one, see How to create your first skill.

How an agent loads a skill

Antigravity loads skills in three stages, often called progressive disclosure:

1. Discovery. At the start of a conversation the agent sees only the name and description of each available skill.

2. Activation. If a skill looks relevant to your request, the agent reads the full SKILL.md.

3. Execution. The agent follows the instructions, and runs any scripts or opens any resources the skill points to.

This keeps the agent's working context small. You can have many skills installed without every one of them being read into every conversation. You can also mention a skill by name in your prompt to make sure it is used instead of waiting for the agent to decide.

Where the folder lives decides who can use it:

| Scope | Location |

|---|---|

| Project | .agents/skills// |

| Global (App / IDE) | ~/.gemini/config/skills// |

| Global (CLI) | ~/.gemini/antigravity-cli/skills/ |

Step-by-step installation is covered in how to install skills in Antigravity.

Not only Antigravity

The SKILL.md format is not unique to Google's tool. Anthropic's documentation for Claude Code says: "Create a SKILL.md file with instructions, and Claude adds it to its toolkit." Its agent skills overview describes the same layered loading: until a skill is triggered, only its name and description occupy context, and the body of SKILL.md is read when it triggers. Folder locations and supported frontmatter fields differ between tools, so check each tool's own documentation before copying a skill from one to another.

Skills vs rules vs MCP servers

These three are easy to confuse because all of them change what your agent does. They solve different problems.

|---|---|---|---|

An MCP server gives the agent a new ability, such as querying a service, that instructions alone cannot provide. They combine well: a skill can tell the agent which MCP tools to use and in what order. For more detail see Antigravity rules and Antigravity MCP servers.

What kinds of skills exist

Skills range from a single page of guidance to a package with bundled code.

Instruction-only checklists. SEO Audit is one Markdown file that sets the agent up as an SEO diagnostic specialist, with a scope gate, checklists and a required format for each finding. Our review found no scripts, no network calls and no keys in it.

Design guidance. UI UX Pro Max gives the agent a local, searchable database of UI and UX guidance plus a Python search script and design-system generator. Its core scripts read bundled data, but the plugin and CLI also install companion skills whose optional scripts call external AI image APIs with your keys or run npx.

Creative output. Canvas Design is Anthropic's skill for posters and static visual pieces. It is instructions plus bundled fonts, with no script and no code.

Browser control. Agent Browser teaches the agent to drive a real Chrome through a command-line tool from Vercel Labs. The skill itself is a small stub, but it hands the agent control of the browser, so the tool behind it matters more than the Markdown.

Web search. Tavily Search tells the agent to run searches through the Tavily CLI. The search happens on Tavily's servers, so every query leaves your machine.

Risks: read a skill before you install it

A skill is text that your agent will follow, and sometimes code it will run. That makes it worth reading the way you would read a script from the internet. Check these things first:

Scripts. Open everything in scripts/. Look for shell commands, curl | bash installers and package installs that run on setup.

Network calls. Does the skill send your prompts, files or code to a third-party service? Search results, image generation and "cloud" helpers usually do.

API keys. Find where keys are read from and what the skill does with them. A skill should not need more credentials than its job requires.

File writes. Check which paths the skill or its scripts write to, and whether they overwrite existing files.

Instruction-only skills carry a smaller attack surface than skills with scripts, but they still steer an agent that can touch your files. We describe our checks in our review method, and How we security-audit skills walks through the checks in more detail. Neither one replaces reading the skill yourself; a review reflects the commit we looked at, and upstream repositories change.

Where to find skills

This directory lists skills from public upstream repositories. Each skill page links to its source, shows install steps and records what we checked in the code and when. To start browsing, see the awesome Antigravity skills list or the full skill directory. We are an independent directory and are not affiliated with Google or the authors of the skills we list.

FAQ

Do skills need code?

No. A skill needs only a SKILL.md file. Scripts, examples and resources are optional, and many skills are pure instructions.

Will the agent always use an installed skill?

Not necessarily. The agent decides from the description whether a skill is relevant. If you want one used, mention it by name in your request.

Can I use the same skill in Antigravity and Claude Code?

Both tools read SKILL.md folders, so a skill is often portable in principle. In practice the folder locations and some frontmatter details differ, so read each tool's documentation and check that the skill works as written.

Are skills safe to install?

It depends on the skill. Text-only skills are lower risk than ones that run scripts, call external services or use API keys. Read the files first and use the checks above.