AI coding agents are getting good at following instructions. But every time you start a new session, you end up re-explaining the same workflows. How to review a PR. How to run a security audit. How to write content in a specific voice. Agent skills solve that problem.

A skill is a packaged, reusable set of instructions that an agent can execute on demand. Think of it as a saved workflow that your agent already knows how to run, without you having to spell it out every time.

The Core Idea

An agent skill is a markdown file (usually a CLAUDE.md or similar config file) that contains structured instructions for a specific task. When you install a skill, it becomes part of your agent’s working context. The agent reads the instructions and follows them when you trigger the skill, either by name or by matching a trigger condition.

Skills are not plugins in the traditional sense. They do not add new APIs or runtime dependencies. They work within the existing capabilities of your agent, directing it to use its tools in a specific, repeatable way.

Here is a concrete example. The Code Reviewer skill tells Claude Code how to perform a structured code review. Instead of writing “review this PR, check for bugs, security issues, and performance problems” every time, you install the skill once. Then you type /review and the agent runs the full workflow: reads the diff, checks for bugs, flags security concerns, evaluates readability, and returns a structured report.

That is the pattern. Package the instructions once. Reuse them forever.

How Skills Work Under the Hood

Most agent skills follow a simple structure:

  1. Trigger definition. What activates the skill. This could be a slash command (/review), a keyword match (“review this PR”), or an explicit invocation.

  2. Context and constraints. What the skill needs to know. File patterns to read, directories to scan, APIs to call, formatting rules to follow.

  3. Execution steps. The actual workflow, written as instructions the agent follows sequentially. Read these files, run this analysis, format the output this way.

  4. Output format. What the result should look like. A report, a diff, a commit message, a summary.

When you install a skill into Claude Code, the instructions get loaded into the agent’s context through the project’s CLAUDE.md file or a dedicated skills directory. The agent does not need a special runtime. It reads the instructions and uses its existing tools (file reading, code editing, shell commands, web search) to execute them.

This is a different thing than MCP servers, which add new tool capabilities through a protocol layer. Skills operate within the tools the agent already has. They are about how the agent uses those tools, not what tools are available.

What Makes Skills Useful

Three things set skills apart from just writing good prompts:

Reproducibility. A prompt you typed last Thursday is gone. A skill persists in your project config. Every session starts with the same capabilities. Every team member gets the same workflow.

Composability. Skills can be combined. A content pipeline might chain a research skill into a writing skill into an editing skill. Each one handles its part, and the output flows to the next step.

Shareability. Skills live in git repos. You can install one from GitHub with a single command, fork it to customize, or publish your own. The AgentNDX skills directory currently tracks over 90 skills across 12 categories, including engineering, security, content creation, and workflow automation.

The Skill Ecosystem Today

Agent skills are still a young category, but the ecosystem is growing fast. Here is what it looks like right now:

By category, the biggest clusters are engineering (code review, testing, deployment), security (vulnerability scanning, dependency audits), and content (writing, research, editing). Newer categories like memory (persistent context management) and orchestration (multi-agent routing) are emerging as agents take on more complex workflows.

By agent compatibility, most skills target Claude Code today. But the format is not locked to one platform. Skills tagged as “universal” work across Claude Code, Codex, Cursor, Gemini CLI, and other agents that support markdown-based configuration. Cross-agent compatibility is where the format gets interesting, and it is moving fast.

By source, you will find official skills from Anthropic (like the Web App Tester and MCP Server Generator), vendor-maintained skills from companies like HashiCorp (Terraform, Vault), and a growing library of community-built skills covering everything from SQL generation to product spec writing.

Skills vs. MCP Servers

This is the question developers ask most often, so here is the short version:

MCP servers add new capabilities. They connect your agent to external systems: databases, APIs, cloud services, browsers. An MCP server for Stripe lets your agent read payment data. Without it, the agent has no access.

Skills direct existing capabilities. They tell your agent how to use the tools it already has. A code review skill does not add new tools. It tells the agent to read the diff, check specific patterns, and format its findings in a specific way.

You often want both. An MCP server gives your agent access to your GitHub repos. A PR review skill tells it how to actually review the code it finds there.

Getting Started

Installing a skill typically takes one command. For Claude Code:

gh skill install anthropics/skills/web-app-tester

This pulls the skill’s instructions into your project. From that point on, the agent knows how to run it.

To browse what is available, the AgentNDX skills directory lists skills by category, compatible agent, and install method. Each entry includes the GitHub source, description, and use cases so you can evaluate before installing.

You can also write your own. A skill is just a markdown file with structured instructions. If you have a workflow you run repeatedly, packaging it as a skill saves time and makes it shareable. The guide to writing your own skill walks through the process step by step.

FAQ

Q: Do skills require any special setup or dependencies?

A: No. Skills are instruction files, not software packages. They work within your agent’s existing tool set. If your agent can read files, run shell commands, and edit code, it can run most skills out of the box.

Q: Can I use skills with agents other than Claude Code?

A: Yes, if the agent supports markdown-based configuration. Many skills in the directory are tagged “universal” and work across Claude Code, Codex, Cursor, and Gemini CLI. Check the compatible agents field before installing.

Q: How are skills different from custom system prompts?

A: System prompts are session-level. They disappear when the session ends. Skills are project-level. They persist in your project config, travel with your repo, and work the same way for every team member. Skills also tend to be more structured, with explicit triggers, steps, and output formats.

Q: Where do I find skills to install?

A: The AgentNDX skills directory tracks over 90 verified skills across 12 categories. You can also find skills on GitHub by searching for repos with CLAUDE.md files that define reusable workflows.