Topic · A1
Google Antigravity Rules and Agent Skills: The Setup Guide
How to configure rules files, AGENTS.md, and agent skills for Google Antigravity — the agent-first IDE that reads the same SKILL.md format as Claude Code and Codex.
The RuleSell Editorial Team
RuleSell's editors review every AI-dev-tooling guide against primary sources — GitHub issues, official docs, and maintainer statements — before publishing.
Reviewed May 30, 2026
# Google Antigravity Rules and Agent Skills: The Setup Guide Google Antigravity is not a code editor with AI bolted on. It's designed as an agent-first development environment — the agent drives; the IDE is the scaffolding. That changes how you configure it. This page covers what you need: how AGENTS.md works in Antigravity, how to bring your existing SKILL.md skills across, and what's different from Claude Code and Cursor.
What Antigravity is
Google announced Antigravity on November 18, 2025. By May 2026 it had Google Developers Blog coverage documenting a second major iteration. It runs Gemini 3.1 Pro by default with Claude and open-source model options. Public preview is free. The key distinction from Cursor or Windsurf: Antigravity's model treats the agent as the primary executor, with the IDE providing the tool surface. A session starts with the agent reading project context (your AGENTS.md), then planning and executing tasks across multiple files and terminals. The human reviews and approves, but the default flow is agent-first. VoltAgent's awesome-agent-skills (23.6k stars) lists Antigravity in its explicit compatibility matrix alongside Claude Code, Codex, Gemini CLI, Cursor, Copilot, OpenCode, and Windsurf. The listed skill count across this ecosystem: 1,424+.The two config layers
Antigravity config separates into two layers: 1. AGENTS.md — the portable project context layer AGENTS.md at your project root is the cross-tool context file that every major agent tool now reads. It carries your stack, build commands, conventions, and constraints. Antigravity reads it at session start. Everything you'd put in a CLAUDE.md or.cursorrules file belongs here if it applies across tools.
See /topic/agents-md for the full AGENTS.md reference and the symlink pattern for Claude Code compatibility.
2. Antigravity-specific rules
For behavior specific to Antigravity — model routing, Gemini-specific prompt patterns, in-IDE UI preferences — use Antigravity's own config layer. At publication time, this is a JSON/YAML config at a Google-documented path; check the official docs as the config format is still evolving in public preview.
Keep the two layers separate. Putting Antigravity-specific instructions in AGENTS.md means every other tool that reads AGENTS.md sees them. Putting cross-tool conventions in an Antigravity-only file means Claude Code and Codex miss them.
Bringing skills to Antigravity
Antigravity reads skills from.agents/skills — the same neutral path that Codex and Goose use. If your skills live in .claude/skills, mirror or symlink them:
``bash
# From your project root
ln -s .claude/skills .agents/skills
`
Or maintain the canonical source in .agents/skills and symlink the other direction for Claude Code:
`bash
ln -s .agents/skills .claude/skills
`
The SKILL.md format works without modification as long as the skill meets the portability criteria below. Antigravity's implicit activation uses the description field, same as Claude Code.
Writing rules and skills that work in Antigravity
Three portability rules matter more in Antigravity than in other tools, because Antigravity may run the skill under Gemini 3.1 Pro rather than Claude.
Keep instructions model-agnostic. Don't write SKILL.md instructions that assume Claude's specific output format, token limits, or reasoning style. Instructions written as "do X, then do Y, return Z format" travel to any model. Instructions written as "as Claude, you will..." don't.
Don't rely on Claude-specific allowed-tools semantics. The allowed-tools field in SKILL.md is a Claude Code-specific execution permission. Antigravity's tool-access model is its own. Write skill logic in instructions; let the host agent decide how to execute it.
Test the description trigger in both runtimes. Claude Code's implicit skill activation and Antigravity's are both description-based, but the matching is implemented separately. A description that fires reliably in Claude Code is likely to fire in Antigravity too — but test it. The /topic/skill-description-engineering rules for precision and trigger words apply equally in both.
Configuring AGENTS.md for Antigravity sessions
If you're using Antigravity as your primary agent and writing AGENTS.md from scratch, include the following that Antigravity's agent-first model particularly benefits from:
Explicit task decomposition hints. Antigravity's planner reads AGENTS.md for context on how to break down work. A section like "Before implementing any feature, check the spec file at docs/specs/" gives the agent a reference point before it starts planning.
Clear deny rules. The agent-first model makes deny rules more important than in ask-mode tools. An Antigravity session may make multiple file edits before pausing for approval. A deny rule that says "never modify database migration files without explicit confirmation" prevents a class of irreversible mistakes.
Tool surface documentation. List the MCP servers and skills the session should use: Use .agents/skills/code-review for any PR review task. MCP config at .mcp.json. The agent reads this as a routing instruction.
A minimal AGENTS.md for an Antigravity session:
`markdown
# Stack
Node 22, TypeScript, Next.js 16, Prisma, Postgres.
# Commands
dev: npm run dev
build: npm run build
test: npm test
lint: npm run lint
# Conventions
- All API routes in src/app/api/
- Use Server Components by default; Client Components only when required
- Prisma schema is the source of truth for DB shape — never modify migrations directly
# Skills
Use .agents/skills/code-review for any review task.
Use .agents/skills/spec-driven-development before implementing a new feature.
# Deny rules
- Never commit to main directly
- Never modify prisma/migrations/ without user confirmation
- Never delete files; propose deletions for approval
``
What to read next
- /topic/agents-md — full AGENTS.md reference and cross-tool symlink patterns
- /topic/agent-skills-cross-tool — verified support per tool including Antigravity
- /topic/run-claude-skills-anywhere — Skillz and Mother MCP for cross-tool skill delivery
- /topic/skill-description-engineering — trigger precision for multi-model environments
- /for-tool/antigravity — the RuleSell catalog filtered for Antigravity-compatible assets
Sources
- Google Developers Blog. "Build with Google Antigravity: Our New Agentic Development Platform". May 20, 2026 coverage of the second iteration. Original announcement November 18, 2025.
- VoltAgent. awesome-agent-skills. 23.6k-star catalog listing Antigravity in its explicit cross-tool compatibility matrix alongside Claude Code, Codex, Gemini CLI, Cursor, Copilot, OpenCode, and Windsurf.
- agents.md. AGENTS.md reference. The cross-tool context file standard read by Antigravity and 20+ other tools.
- Anthropic. Extend Claude with skills. Claude Code skill discovery model (context for what Antigravity mirrors and where it diverges).
Frequently asked
- What is Google Antigravity?
- Antigravity is Google's agent-first development environment, announced November 18, 2025 and covered in a Google Developers Blog post on May 20, 2026. It runs on Gemini 3.1 Pro by default but also supports Claude Sonnet/Opus and open-source models. It's listed in VoltAgent's awesome-agent-skills compatibility catalog alongside Claude Code, Codex, Cursor, Copilot, Gemini CLI, Windsurf, and OpenCode. As of mid-2026 it is in free public preview. ([source: developers.googleblog.com](https://developers.googleblog.com/build-with-google-antigravity-our-new-agentic-development-platform/))
- Does Antigravity read AGENTS.md?
- Yes. Like most post-2025 agent tools, Antigravity reads AGENTS.md at the project root for context — stack, commands, conventions, deny rules. It also reads a tool-specific rules layer (analogous to .cursorrules in Cursor) for Antigravity-specific configuration. The general pattern is: AGENTS.md for the portable cross-tool layer, Antigravity-specific files for anything that uses Gemini-specific behavior.
- Does Antigravity support SKILL.md?
- Yes, based on its inclusion in the VoltAgent awesome-agent-skills compatibility list, which documents support across Claude Code, Codex, Antigravity, Gemini CLI, Cursor, Copilot, OpenCode, and Windsurf. The skill discovery path follows the .agents/skills convention that Codex and Goose also use. Skills authored with a neutral SKILL.md — minimal frontmatter, portable instructions, no Claude-specific bash hooks — should activate in Antigravity without modification.
- Can I use Claude Code skills in Antigravity?
- Skills with portable instruction text and no Claude-specific execution hooks work fine. Skills that depend on Claude Code's allowed-tools or bash hook model will present their instruction text in Antigravity but won't trigger the same execution semantics. The practical rule: write the logic in instructions, not in scripts, if you want the skill to work identically in both.
- What models does Antigravity support?
- Gemini 3.1 Pro is the default. The system also supports Claude Sonnet and Claude Opus (via Anthropic's API) and selected open-source models. This means rules and skills that reference Claude-specific output formats will work in Antigravity sessions using Claude, but may behave differently in a Gemini session if the instruction assumes Claude's specific response patterns.