AI Agent Skills: Claude, Cursor & Antigravity Compared

Skills are structured knowledge bundles that persist across sessions. Instead of re-explaining your stack, conventions, and preferences every time, you define them once and the agent references them when triggers match.

A rule file tells the model what you want. A skill file tells it how to verify it did the right thing.

The Three Contenders

Right now, three major platforms support skills in meaningfully different ways:

1. Claude Skills (Anthropic)

Anthropic open-sourced their skills framework on GitHub. Skills are markdown files with YAML frontmatter defining triggers, descriptions, and structured instructions.

1# SKILL.md
2---
3name: nextjs-patterns
4description: Next.js App Router patterns and conventions
5triggers:
6  - "next.js"
7  - "app router"
8  - "server components"
9---
10
11## Core Patterns
12
13- Always use Server Components by default
14- Add 'use client' only when necessary
15- Prefer Server Actions over API routes
16...

Pros: Open source, well-documented, integrates with Claude Projects.
Cons: Requires Claude API or Pro subscription.

2. Cursor Agent Skills

Cursor's approach is file-system based. You drop SKILL.md files in .agent/skills/ directories, and the agent discovers them automatically.

1.agent/
2└── skills/
3    ├── nextjs-routing/
4    │   ├── SKILL.md
5    │   └── examples/
6    └── testing/
7        ├── SKILL.md
8        └── scripts/

Pros: Lives in your repo, version controlled, discoverable.
Cons: Cursor-specific, no cross-platform portability.

3. Google Antigravity Skills

Antigravity (the codename for Google's advanced agentic coding assistant) uses a similar file-based approach but with deeper integration into Gemini's reasoning capabilities.

1# .agent/skills/storybook/SKILL.md
2---
3name: storybook-mastery
4description: Generate production-quality Storybook stories
5---
6
7## Protocol Layers
8
91. Configuration enforcement
102. Framework adapters (Next.js, Radix)
113. Interaction testing requirements
124. Self-correction checklist

Pros: Deep reasoning integration, multi-step verification, access to Google Search for grounding.
Cons: Newer, less documentation available.

Converting a Rule to a Skill

Here's the pattern I use to convert a .cursorrules file (or any system prompt) into a proper skill:

Step 1: Extract the Core Behavior

Most rules are a mix of "do this" instructions and "don't do that" constraints. Separate them:

1# Before (flat rule)
2"Always use TypeScript. Never use any. Always add tests."
3
4# After (structured skill)
5## Requirements
6- Use TypeScript for all new files
7- Add unit tests for utility functions
8
9## Constraints  
10- Never use `any` type
11- Never skip error handling

Step 2: Add Contextual Triggers

Skills work best when they activate automatically based on context:

1---
2triggers:
3  - file:*.stories.tsx
4  - prompt:"storybook"
5  - prompt:"component story"
6---

Step 3: Include Examples

The difference between a mediocre skill and a great one? Concrete examples. AI learns from patterns:

1// Correct
2export const Default: Story = {
3  args: { variant: 'default' },
4  play: async ({ canvasElement }) => {
5    // interaction test
6  }
7};
8
9// Wrong
10export const Default = () => <Button>Click</Button>;

Step 4: Add Self-Correction

This is the secret sauce. Force the AI to verify its own output:

1## Pre-Output Checklist
2
3Before returning code, verify:
41. [ ] Did I use the @/ alias?
52. [ ] Did I add the 'autodocs' tag?
63. [ ] Did I wrap Radix components in providers?
74. [ ] Did I include a play function for interactions?

Which Should You Use?

My recommendation? Use what your team already has. If you're in VS Code with Cursor, go Cursor. If you're deep in the Google ecosystem, Antigravity is worth exploring.