The definitive guide to understanding agent skills — what they are, how they work, why developers use them, and how to get started with the open SKILL.md standard. Agent skills explained from fundamentals to advanced usage.
Whether you are asking "what is agent skills" for the first time or looking to deepen your expertise, this guide covers everything you need to know about the open SKILL.md standard powering 14+ AI coding tools.
So what are agent skills, exactly? Agent skills are modular, reusable packages of instructions, workflows, and domain knowledge that give AI coding agents specialized capabilities. Each one is stored as a directory containing a required SKILL.md file with YAML frontmatter (metadata such as name, description, and tags) followed by markdown instructions that tell the agent how to behave when the skill is activated. This approach transforms AI coding assistants from general-purpose tools into domain-specific experts.
Unlike one-off prompts that vanish after a single session, agent skills persist across every conversation. Once installed, the agent recognizes when a skill is relevant and loads it automatically — no manual activation required. This persistence is what makes them fundamentally different from ad-hoc prompting. You encode your best practices once, and they apply reliably every time.
They can range from simple single-file instructions to comprehensive collections that include multiple sub-skills, reference documentation, templates, and even helper scripts. The format is intentionally text-based and human-readable — anyone who can write markdown can create one. This low barrier to entry has fueled rapid ecosystem growth, with hundreds of thousands now available across community repositories.
Under the hood, agent skills follow a lazy-loading architecture designed for token efficiency. When a session begins, the AI agent reads only the frontmatter from each installed skill — typically just the name, description, and tags. This metadata index costs very few tokens, so you can have dozens installed without any noticeable performance impact.
When the agent determines that a particular skill is relevant to the current task (based on the metadata match), it loads the full instruction body on demand. This on-demand activation means the skill consumes tokens only when actually needed. The alternative — stuffing all instructions into a system prompt — would quickly exhaust the context window.
The SKILL.md format is the heart of how agent skills work. A typical file starts with YAML frontmatter specifying the skill name, a short description, version, and optional tags. Below the frontmatter, markdown sections contain detailed instructions, rules, examples, and workflows. Some advanced skills also include companion files — templates, configuration snippets, or reference documentation — in the same directory.
Because they are plain text files, skills integrate naturally with version control. The agent skills md format (SKILL.md) is often referred to as the agent skills standard — a single, open specification that every compatible tool understands. Teams commit their agent skills md files alongside source code, track changes through pull requests, and review updates just like any other project file. This makes them the first truly developer-native approach to customizing AI coding assistants.
The concept of agent skills originated at Anthropic, who created the SKILL.md format for Claude Code in 2025. The idea addressed a gap in the AI coding workflow: developers were repeatedly typing the same instructions across sessions, and there was no standardized way to package and share those instructions. The SKILL.md format solved this by introducing a file-based approach that the agent could discover and load automatically.
The format quickly gained traction beyond Claude Code. OpenAI adopted the standard for Codex CLI in late 2025, recognizing the value of a shared, open format. Shortly after, GitHub integrated agent skills into Copilot, and tools like Cursor, Windsurf, and Gemini CLI followed. By early 2026, the SKILL.md standard had been adopted by 14+ AI coding tools — an unprecedented level of cross-tool interoperability in the AI development ecosystem.
Today, the format is recognized as a foundational layer in the AI coding stack. They sit alongside Model Context Protocol (MCP) as one of the two primary ways to extend AI agent capabilities. While MCP connects agents to external systems, skills provide the procedural knowledge that tells agents how to accomplish tasks. The rapid adoption across competing platforms demonstrates that the industry views them as essential infrastructure, not a proprietary feature.
Understanding what are agent skills becomes clearer when you compare them to traditional prompting. Sometimes called agentic skills because they empower agentic workflows, they represent a significant improvement over ad-hoc prompts.
| Aspect | Agent Skills | Ad-Hoc Prompts |
|---|---|---|
| Persistence | Persist across sessions — install once, use forever | Lost after each session, must re-type |
| Reusability | Fully reusable across projects and teams | Copy-paste between sessions, prone to drift |
| Shareability | Share via Git, ZIP, or marketplace | Shared informally via docs or chat |
| Version Control | Git-tracked with full history | No versioning |
| Token Efficiency | Lazy-loaded on demand | Consumes tokens every time |
| Cross-Platform | Works in 14+ tools via SKILL.md | Tool-specific, not portable |
In short, agent skills turn ephemeral prompting knowledge into durable, portable assets. Teams that adopt them report saving 2-3 hours per developer per week by eliminating repetitive instruction writing.
Agent skills and MCP are complementary technologies, not competitors. Skills provide procedural knowledge — they teach the agent how to perform tasks through step-by-step instructions, coding standards, and workflow definitions. MCP provides tool connectivity — it gives the agent access to external systems like databases, APIs, and services.
Think of it this way: skills are the playbook, and MCP is the toolkit. A deployment skill might define how to deploy an application (the workflow, safety checks, and rollback procedures), while an MCP server provides the actual connection to your cloud infrastructure. Most teams use both together for maximum effectiveness.
A key practical difference is setup complexity. Agent skills require no server — you simply drop a SKILL.md file into the right directory. MCP requires running a server process and configuring JSON-RPC connections. Skills are also more portable: one file works across all compatible platforms, while MCP server implementations can vary by tool.
For a comprehensive analysis of how they compare to MCP and why many developers consider them equally transformative, read our detailed article: Agent Skills: A Bigger Deal Than MCP?
Traditional IDE plugins and skills solve related problems but in fundamentally different ways. IDE plugins are compiled extensions that hook into a specific editor's API — they require programming knowledge to create, are tied to a single platform, and need updating whenever the host editor changes its API. Skills, by contrast, are plain text markdown files that work across every tool supporting the SKILL.md standard.
This text-based simplicity is a core strength. Anyone who can write markdown can create an agent skill — no SDK, no build step, no platform-specific code. They are also dramatically lighter: a typical skill is a few kilobytes of text, while a plugin can be megabytes of compiled code. Because they operate through natural language instructions rather than API bindings, they are inherently more flexible and easier to update.
The tradeoff is that they cannot access low-level editor APIs (syntax highlighting, custom UI panels, keybindings). But for the vast majority of AI-assisted development workflows — code generation, testing, review, documentation, deployment — skills provide everything you need without the overhead of traditional plugin development. As the ecosystem matures, this lightweight approach is proving to be a major advantage.
The agent skills ecosystem covers a wide range of use cases across the software development lifecycle. Here are the primary categories you will find when exploring the agent skills directory.
Now that you understand what are agent skills, here is a step-by-step guide to start using them in your daily workflow.
Start by exploring the agent skills marketplace to discover skills that match your workflow. Filter by category, platform, or use case. Each listing includes a description, preview, compatibility information, and download button.
Download the skills you want as a ZIP file. For Claude Code, extract to ~/.claude/skills/. For Codex CLI, place them in .codex/skills/. For other tools, copy the SKILL.md file to the tool's skills directory. The agent discovers and indexes them automatically on the next session start.
Once installed, you do not need to explicitly activate them. Simply work as you normally would. When your task matches an installed skill, the agent loads it automatically and applies the encoded instructions. You will notice immediately more consistent, higher-quality output aligned with the skill's defined best practices.
Beyond the marketplace, thousands of community-created skills are available on GitHub. Visit our agent skills GitHub page for a curated list of the best repositories. You can also search GitHub directly for SKILL.md files to discover skills built by other developers.
Once comfortable with existing ones, start creating your own agent skills. Any workflow you repeat frequently is a candidate. Create a directory with a SKILL.md file, define the frontmatter, and write clear markdown instructions. Share them with your team or publish to the community through the agent skills marketplace.
Agent skills are modular instruction packages stored as SKILL.md files that give AI coding agents specialized capabilities. When you install them, the agent reads only the name and description at session start. Full instructions load on demand when relevant to your task — a lazy-loading design that keeps token usage efficient even with dozens installed.
Agent skills use the open SKILL.md standard, which is supported by 14+ AI coding tools including Claude Code, Codex CLI, ChatGPT, GitHub Copilot, Cursor, Windsurf, and Gemini CLI. One SKILL.md file works across all compatible platforms with no conversion needed. This cross-platform compatibility is a core advantage over proprietary plugin systems.
Yes. All agent skills on Claude Skills Hub are free and open source, published under permissive licenses like MIT and Apache 2.0. You can use them in personal and commercial projects at no cost. The SKILL.md standard itself is open, so anyone can create, share, and distribute them freely.
Agent skills excel at encoding repeatable workflows, coding standards, testing patterns, security checklists, and domain-specific knowledge. They are best suited for tasks you perform frequently — code review, test generation, API design, documentation, and deployment automation. Any workflow you explain to an AI agent more than once is a candidate for becoming a skill.
Now that you know what are agent skills, dive deeper into the ecosystem. Browse, search, and discover them through these resources.
Browse the complete collection with filtering, categories, and platform compatibility.
The largest curated marketplace for AI skills. Download, preview, and install with one click.
Find open-source skills on GitHub. Explore repositories, contribute, and publish your own.
Complete directory organized by category, platform, and use case. Find the right skills for any workflow.
Subscribe to get the latest Claude Skills, tutorials, and community highlights delivered to your inbox.
We respect your privacy. Unsubscribe at any time.