Google Antigravity Agent Skills: Architecture, Formats, and Portability

What Is the Google Antigravity Skill Architecture?

Google Antigravity skills are specialized, self-contained capability packages that provide autonomous agents with domain-specific procedural instructions, tool configurations, reference documentation, and execution scripts. Rather than relying on monolithic system prompts that exhaust token capacity, Google Antigravity organizes agent capabilities into discrete directories governed by a central SKILL.md manifest.

As documented in the official Anthropic Claude Code Documentation: “Skills are modular packages containing a SKILL.md file and optional supporting files that provide reusable instructions, scripts, and context to Claude Code.” Google Antigravity adopts this standardized SKILL.md structure, expanding it to orchestrate complex multi-agent architectures, delegated subagent execution via invoke_subagent, and custom Model Context Protocol (MCP) integrations.

In the Antigravity runtime, skills operate as on-demand capabilities. When the agent initializes a session, it reads only the lightweight 35-to-50 token YAML frontmatter header containing the skill's name and description. The full procedural body and reference templates load into context only when the agent's planner identifies a task matching the skill's operational domain, achieving a 95% reduction in baseline context overhead.

Where Are Google Antigravity Skills Stored and Discovered?

Google Antigravity resolves agent skills across three distinct filesystem tiers: system builtins, user-installed plugin packages, and workspace-scoped project repositories. This multi-tiered directory structure ensures that global developer habits remain available across all projects while project-specific engineering rules take precedence when collaborating in team repositories.

The three standard filesystem discovery paths in Google Antigravity include:

# 1. Built-in Core Antigravity Skills
~/.gemini/antigravity/builtin/skills/<skill-name>/SKILL.md

# 2. Plugin & Extension Skills (Ecosystem Tools)
~/.gemini/config/plugins/<plugin-name>/skills/<skill-name>/SKILL.md

# 3. Workspace / Project-Level Skills (Repository Scope)
<workspace-root>/.gemini/skills/<skill-name>/SKILL.md
<workspace-root>/skills/<skill-name>/SKILL.md

When an autonomous agent processes a prompt, the Antigravity planner evaluates available skills in reverse hierarchy: project-level skills override plugin-defined skills, which in turn override global builtin defaults. This resolution order allows teams to standardize on global engineering conventions while overriding specific migration or deployment workflows within individual microservices.

The Sprawl Moment: When Your Antigravity Skills Fail to Follow You

A senior staff engineer spends three weeks perfecting an advanced database migration and rollback skill inside Google Antigravity on their primary macOS development machine. The skill contains intricate Zod schema assertions, automated dry-run verifications, and custom MCP query hooks. On Wednesday morning, the engineer opens a remote Linux cloud workstation to handle an urgent production outage using Claude Code and OpenAI Codex. When they invoke the migration command, the agent defaults to generic, unconstrained SQL generation because the customized Antigravity skill directory only exists on the local MacBook. Grepping through disconnected dotfile backups and searching ancient Slack threads burns 45 minutes of critical incident response time, forcing the engineer to manually re-author the validation logic from memory.

Prompttly is a skill manager for AI agents — one library for your skills and prompts that syncs into Claude Code, Codex, ChatGPT, and Claude and is one hotkey away on your Mac, so your setup follows you across every machine, repo, and tool.

How Do Antigravity Skills Compare to Claude Code, Codex, and Cursor Rules?

Google Antigravity skills, Claude Code skills, OpenAI Codex skills, and Cursor rules share the common goal of guiding AI behavior, but they differ significantly in storage paths, lazy-loading mechanics, and multi-agent invocation capabilities. Understanding these structural boundaries prevents configuration errors and instruction collisions.

The comparison table below details the architectural specifications across each primary AI coding platform:

To learn more about how instruction files differ across tool environments, read our technical guides on SKILL.md vs CLAUDE.md vs AGENTS.md, OpenAI Codex skills, and Cursor rules vs Claude skills.

Multi-File Skill Anatomy: SKILL.md, Scripts, and References

A production-grade Google Antigravity skill is packaged as a modular folder containing 4 standard structural components: the root SKILL.md manifest, an executable scripts/ directory, a references/ documentation directory, and an examples/ directory. This separation of concerns prevents reference data from bloating the agent's active prompt window.

The canonical directory layout of an advanced Antigravity skill is structured as follows:

database-migration-auditor/
├── SKILL.md                 # Primary entrypoint with YAML frontmatter
├── scripts/                 # Executable validation utilities (Node/Python/Bash)
│   ├── check_schema.py      # Dry-run schema validation script
│   └── verify_indexes.sh    # Index collision checker
├── references/              # Domain specifications & constraints
│   ├── naming_rules.md      # Team naming conventions
│   └── zero_downtime.md     # Safe DDL migration rules
└── examples/                # Golden samples & test scenarios
    ├── safe_migration.sql   # Compliant schema change sample
    └── risky_migration.sql  # Dangerous lock-inducing anti-pattern

The SKILL.md manifest defines the skill metadata and provides clear execution logic. Here is a production-ready example:

---
name: database-migration-auditor
description: Audits SQL migration files for zero-downtime compliance, lock contention risks, and missing index definitions before execution.
---

# Database Migration Auditor

You are an expert PostgreSQL DBA reviewing migration scripts.

## Execution Workflow
1. Parse all SQL statements in the proposed migration file.
2. Execute `python scripts/check_schema.py --file <target_file>` to verify table constraints.
3. Cross-reference table alterations against `references/zero_downtime.md`.
4. If an exclusive table lock is detected without a CONCURRENTLY clause, reject the change.
5. Provide a remediation diff conforming to `examples/safe_migration.sql`.

For a deeper breakdown of multi-file package design, explore our comprehensive guide on multi-file skills and packaging.

How Do You Make Antigravity Skills Portable Across Multiple AI Agents?

Making Google Antigravity skills portable across Claude Code, OpenAI Codex, and Cursor requires isolating platform-agnostic procedural logic from agent-specific tool declarations. When instructions avoid hardcoded CLI paths or proprietary system prompt macros, the same skill package executes reliably across any modern coding assistant.

Follow these 4 core architectural principles to achieve universal agent portability:

1. Standardize on Strict YAML Frontmatter: Always format the top of SKILL.md with standard name and description attributes enclosed in triple dashes. Avoid proprietary frontmatter extensions that cause YAML parsing exceptions in Claude Code or Codex.
2. Use Relative Directory Paths: When referencing helper scripts in scripts/ or documentation in references/, use relative paths from the skill root rather than absolute filesystem paths like /Users/name/.gemini/....
3. Decouple Agent Tool Hooks from Instructions: Keep instructions focused on what steps to take and what output constraints to enforce, rather than hardcoding agent-specific function call names.
4. Maintain an Authoritative Central Library: Instead of manually symlinking directories across ~/.gemini/antigravity/, ~/.claude/skills/, and ~/.codex/skills/, use a dedicated skill manager to mirror packages across all agent directories automatically.

Step-by-Step Migration: Exporting and Syncing Antigravity Skills

Migrating an existing collection of Google Antigravity skills into a synchronized, multi-agent library involves 4 structured steps: auditing local skills, validating frontmatter compliance, configuring agent directory mirrors, and verifying execution across tools.

1. Step 1: Audit Local Antigravity Skills. Inspect ~/.gemini/antigravity/builtin/skills/ and your active project repositories to list all custom SKILL.md definitions.
2. Step 2: Clean and Validate Metadata. Ensure every skill contains an actionable description that clearly outlines trigger conditions. You can use the free Claude Skill Creator or Custom Instructions Generator to generate structured manifests.
3. Step 3: Establish Multi-Agent Filesystem Sync. Configure synchronization to mirror the canonical skill library into ~/.claude/skills for Claude Code and ~/.codex/skills for Codex.
4. Step 4: Verify Multi-Agent Execution. Test invoking the skill in Antigravity, verify slash-command recognition in Claude Code, and test instant retrieval via sub-500ms global Mac hotkeys inside your desktop IDE.

For complete synchronization setup details, review our guide on how to sync AI skills across computers and our breakdown of why AI power users need a skill manager.

Related Resources and Next Steps

Explore the complete library of AI workflow guides on the Prompttly Resources Hub. Read our companion architectural guides on SKILL.md explained, Claude Code skills folder locations, OpenAI Codex skills, and updating AI skills safely.

To stop losing skills across laptops, automatically sync your library into Google Antigravity, Claude Code, and Codex, and summon any prompt in 500ms with a global hotkey, explore Prompttly Plans or download the Prompttly Mac App.

Frequently Asked Questions About Google Antigravity Skills