Skip to content

Introduction to Agent Skills

You just spent twenty minutes explaining to the AI how your team handles deployments. You described the branching strategy, the CI pipeline, the staging environment URL, the Slack channel for release notifications, and the rollback procedure. The AI got it right and deployed successfully. Tomorrow, you open a new session and have to explain everything again. The AI has no memory of yesterday’s conversation.

Skills solve this. A skill is a set of instructions, conventions, and rules that persists across sessions. Instead of re-explaining your deployment process every time, you install a skill once and every future session starts with that knowledge already loaded.

  • Understanding of what skills are and how they differ from MCP servers
  • Setup for the skills CLI across all three tools
  • Your first skill installed and working
  • Overview of the skills marketplace and how to discover useful skills

A skill is a markdown file (or collection of files) that contains instructions for an AI coding agent. When installed, the skill’s content is automatically included in the AI’s context at the start of every session. Skills can contain:

  • Coding conventions — “Always use TypeScript strict mode, prefer interfaces over types for object shapes”
  • Framework patterns — “When building React components, use this file structure and naming convention”
  • Tool configurations — “Here is how to use our internal API, including authentication and rate limits”
  • Workflow instructions — “When deploying, follow this sequence of commands and checks”

Skills are stored in Git repositories and installed via the skills CLI. The marketplace at skills.sh indexes thousands of community-contributed skills.

  1. A skill author creates a repository with markdown files containing instructions and rules for the AI.
  2. You install the skill using npx skills add <owner/repo>. This clones the instructions into your project’s .claude/skills/ directory.
  3. Your AI agent reads the skill at the start of every session, applying the instructions to all subsequent interactions.
  4. Skills persist across sessions. Unlike conversation context, skills remain installed until you remove them.

The most popular skill is find-skills, which teaches the AI how to discover and install other skills from the marketplace. Start here.

Skills integrate with Cursor through the .cursor/rules/ directory. Install via the terminal:

Terminal window
npx skills add vercel-labs/skills

This installs the find-skills skill (among others in the repo) as a rules file that Cursor reads automatically.

The skills.sh marketplace is the central directory for community skills. As of early 2026, it indexes over 200 skills across categories:

CategoryExamples
FrontendReact best practices, Next.js patterns, Vue composition API, Svelte conventions
BackendNode.js patterns, Python best practices, Go conventions, Rust idioms
DevOpsDocker workflows, Kubernetes patterns, Terraform conventions
TestingJest patterns, Playwright best practices, testing strategies
FrameworksAstro, Remix, SvelteKit, Nuxt, Rails, Django
ToolingGit workflows, CI/CD patterns, linting conventions

The marketplace leaderboard gives a sense of what the community finds most useful:

SkillInstallsWhat It Does
vercel-labs/skills (find-skills)154,000+Teaches the AI to search and install skills
vercel-labs/agent-skills (vercel-react-best-practices)107,000+Vercel’s official React conventions, web design guidelines, and composition patterns
anthropics/claude-code65,000+Anthropic’s official Claude Code skills (frontend design, skill development, agent development, hooks)
cloudflare/skills40,000+Cloudflare development skills (Wrangler, Workers, Durable Objects, Agents SDK)
stripe/agent-toolkit35,000+Stripe API integration patterns

Skills and MCP servers are complementary, not competing:

  • Skills teach the AI how to think. They provide instructions, conventions, and patterns that shape the AI’s approach to coding tasks.
  • MCP servers give the AI hands. They provide tools for interacting with external systems — reading files, querying databases, managing pull requests.

A skill might say: “When building React components, always use named exports, create a test file alongside the component, and follow the team’s naming convention.” An MCP server provides the actual capability to create files, run tests, and commit code.

The most effective setups use both: skills to establish conventions and MCP servers to execute actions.

Skills do not load in a new session. Verify the skill files exist in your project’s skills directory (.claude/skills/ for Claude Code, .cursor/rules/ for Cursor). Run npx skills list to see installed skills.

AI ignores skill instructions. If the context window is full from MCP server tool descriptions and conversation history, skills may get truncated. Disconnect unused MCP servers to free up space for skill instructions.

“npx skills” command not found. The skills CLI is distributed as an npm package. Ensure Node.js is installed and npx is available in your PATH.

Conflicting skills. Two skills might give contradictory instructions (one says “use classes,” the other says “use functions”). Review installed skills with npx skills list and remove conflicts.