Antigravity Skills: How to Use, Add, and Create Them

What are Antigravity Skills?

Antigravity Skills are a way to package repeatable instructions for agent work. Instead of pasting the same setup notes into every prompt, you can keep a focused workflow in a named skill and let the agent discover it when the task matches the description. A good skill is not a giant rulebook: it is a small, testable unit that explains when it applies, what it needs to read, what steps to follow, and what a finished result should look like.

The useful distinction is between capability and permission. A skill can teach an agent how to audit a sitemap, run a release checklist, or prepare a component. It does not automatically make every command safe. The agent may still need approval for shell commands, network access, secrets, file deletion, package installation, or MCP tools. Keep the instructions explicit about boundaries so the skill improves consistency without turning a useful shortcut into silent automation.

Search demand around Antigravity Skills also includes GitHub collections, installation questions, and comparisons with rules. Those are related but not identical intents. This page focuses on the practical path from a blank folder to a verified skill, then routes plugin, CLI, MCP, and permission questions to the more appropriate guide.

Where do Antigravity Skills live?

The exact location depends on whether you want the skill available to one project or to your user account. Follow the current official documentation for the product surface you are using, then confirm the active workspace and account before debugging discovery. A common mistake is creating a valid SKILL.md in a directory that the current agent session never scans.

Start with project scope when a skill depends on repository scripts, local conventions, or a particular deployment process. Use user scope for a general workflow that you can test across several repositories. Keep project-specific paths and secrets out of a user-wide skill, and do not copy a community skill into a sensitive repository until you have read every referenced file.

How to create an Antigravity Skill with SKILL.md

Create the smallest useful skill first. Pick one job such as checking a static-site release, reviewing a migration, or preparing a test plan. Give the folder a stable lowercase name, then write a short description that tells the agent what the skill does and when it should be selected. If the description is vague, the skill may never be discovered; if it is too broad, it may be selected for work it cannot safely complete.

Keep the main file readable. Put the workflow, constraints, expected evidence, and stop conditions in SKILL.md. Move long examples or narrow reference material into child files and tell the agent when to read them. A skill that asks the agent to run a command should name the command, explain the expected output, and say what to do when the result is different.

1. Choose one taskWrite the user decision the skill helps with, not a list of every tool in your stack.
2. Create the folderUse a stable name and keep the skill inside the project or user scope you intend to test.
3. Write frontmatterInclude a clear name and description so discovery can match the task naturally.
4. Add the workflowList inputs, ordered steps, evidence to collect, safety boundaries, and the definition of done.
5. Test with a small taskAsk for a dry run or read-only check first, inspect the output, and only then allow write operations.

---
name: sitemap-release-check
description: Check a static sitemap before a release and report missing URLs, bad canonicals, and stale lastmod values.
---

# Sitemap release check

1. Read the project instructions and identify the source sitemap.
2. Inspect the route and the generated HTML before changing files.
3. Run the project's validation command with UTF-8 output.
4. Report evidence and stop before destructive cleanup.

Read references/checklist.md only when the repository uses the extended audit rules.

The example is a structure illustration, not a promise that every release uses the same folder name or command. Match the current Google documentation and your repository conventions before publishing a skill.

How the agent uses a skill

Discovery usually starts with the skill name and description. When a task looks like a match, the agent should read the main instructions and then load only the references needed for that task. This progressive-disclosure pattern keeps the first context small and makes it easier to see which details influenced the answer. It also gives you a natural review point: inspect the skill file before you trust the workflow it proposes.

A useful skill describes its own limits. Say whether it is read-only, which directories are in scope, which commands are safe, what requires confirmation, and what evidence should be saved. If the task changes a repository, require a final diff review and a clean status check. If the skill touches external services, name the official endpoint or source and avoid silently substituting a mirror.

The first run should be boring. Ask the agent to inventory files, explain the proposed plan, or validate a harmless fixture. Treat a surprising command, unexplained network request, or request for a token as a stop signal rather than an invitation to widen permissions.

Skills vs rules, plugins, and MCP

The phrase “Antigravity customization” covers several different surfaces. A new page is useful only when the boundary is clear. Use a skill for a repeatable task procedure, a rule for a persistent instruction that should shape many tasks, a plugin or extension for product capabilities and integrations, and MCP for a tool or data connection. They can work together, but they should not be described as interchangeable.

The related keyword antigravity rules has enough demand for its own future research, while antigravity plugins already maps to the site's Extensions guide. For this page, keep the comparison practical: decide what belongs in the skill file, link to the relevant existing guide, and avoid duplicating a complete MCP or CLI tutorial here.

Install and verify a community skill safely

A GitHub collection can be useful, but a search result is not a trust signal. Before copying a skill, inspect the repository history, license, referenced scripts, package files, network calls, and required credentials. Read the full SKILL.md and any linked reference or script files. A short skill can still ask an agent to run a destructive command if you do not inspect the details.

After installation, verify discovery in the smallest workspace that reproduces the intended task. Start with a read-only request and ask the agent to explain which skill it selected and why. Compare the expected files with the actual diff. If a skill depends on a CLI, plugin, or MCP server, test that dependency separately so a failure is not misdiagnosed as a skill-discovery problem.

1. Read the sourceCheck the publisher, license, commit history, references, scripts, dependencies, and external endpoints.
2. Copy into a test scopeUse a disposable repository or a project-level folder before making the skill user-wide.
3. Run a read-only promptAsk for an inventory, explanation, or dry-run report before allowing edits or shell actions.
4. Inspect evidenceConfirm the selected skill, files read, commands proposed, and output match your intent.
5. Promote carefullyOnly share the skill more broadly after the workflow is predictable and its permissions are documented.

• Do not place API keys, browser session data, or production credentials in a skill file.
• Keep delete, deploy, payment, messaging, and account-administration actions behind manual review.
• Pin or review dependencies when a skill calls a script, package manager, or external service.
• Use version control so a generated change can be inspected and reversed.
• Treat an unfamiliar network request as a reason to stop and investigate.

Version and official-source check

This project is a software download site, so the version gate was completed before publishing the guide. On August 19, 2026, the official Antigravity documentation and download flow were checked for the current product lines. The desktop line recorded by this site is Antigravity 2.0 v2.8.1 with Antigravity IDE v2.5.5; the CLI guide was synchronized to v1.1.14. Skills behavior can change separately from installer versions, so use the official Skills documentation as the final authority.

The download call to action remains an official-page fallback. No permanent binary URL, package size, checksum, or safety-scan result is invented here. Historical v1.13.3 and v2.2.1 pages remain archives and are not presented as the current release.