The Agent Skills Standard: Google Antigravity vs. Claude Code

Summary: The industry is rapidly converging on a standard for "teaching" AI agents: Markdown-based Skills. We dissect the differences between Google Antigravity and Claude Code—two heavyweights using the same SKILL.md format but with vastly different philosophies.

It’s official: Direct instruction via Markdown is the new configuration standard for AI Agents.

Both Google Antigravity and Anthropic's Claude Code have adopted the "Skill" concept—a folder containing a SKILL.md file that teaches an agent a specific workflow. But under the hood, their implementations reveal distinct philosophies about how humans should control AI autonomy.

The Common DNA

Before we diverge, let's look at what's identical. Both platforms have agreed on:

1. File Format: SKILL.md as the source of truth.
2. Discovery: YAML frontmatter (name, description) used by the LLM to decide when to trigger the skill.
3. Scope: Support for both "Project" (repo-level) and "Global" (user-level) skills.

This is a massive win for developers. A single standard means you might soon write a "Deploy to AWS" skill that works across any agentic IDE.

The Divergence: Complexity vs. Simplicity

Where they differ is in how they handle power.

1. Google Antigravity: The "Keep It Simple" Approach

Google’s implementation is ruthlessly simple. It treats the Skill primarily as a context wrapper around standard scripts.

• Philosophy: "The Agent provides the intelligence; you provide the script."
• Tooling: No special permissions config. If the agent has terminal access, it can run your script.
• Structure:
  markdown

  Copy

  # Deploy to Stage
  run `./scripts/deploy.sh`
  

Pros:

• Zero friction to start.
• "Scripts as Black Boxes" keeps the context window clean.

Cons:

• Less granular control over what the agent can do inside that skill.

2. Claude Code: The "Granular Control" Approach

Anthropic puts the configuration inside the skill definition, offering fine-grained control for complex orchestrations.

• Philosophy: "The Skill is a configurable program."
• Tooling: Explicit allowed-tools in frontmatter. You can ban file writing for a "Read-Only Analysis" skill.
  yaml

  Copy

  allowed-tools: Read, Grep, Glob # No Write access!
  

• Lifecycle Hooks: Claude allows PreToolUse and PostToolUse hooks defined directly in YAML.
• Sub-Agents: You can fork the context (context: fork) to spawn a sub-agent specifically for that skill.

Pros:

• Enterprise-grade security controls (e.g., preventing accidental writes).
• Powerful orchestration capabilities with hooks and sub-agents.

Cons:

• Higher learning curve.
• More boilerplate YAML.

Feature Showdown

Verdict: Which approach is right for you?

Choose Google Antigravity if...

You want to "Get Shipping." You have a folder of bash scripts and you want an agent to run them intelligently. You prefer your logic to live in projects/scripts/*.sh rather than inside prompt definitions.

Choose Claude Code if...

You are building "Agentic Architectures." You need to chain multiple agents, restrict permissions for safety (e.g., a "Safe Refactor" skill that can't delete files), or distribute complex flows via Plugins.

The Future is Markdown

Regardless of your camp, the news is good: We are moving away from proprietary JSON configurations and towards natural language SKILL.md files. This is the "HTML" moment for Agents—a common markup language that will define how we work with AI for the next decade.