Claude Skills are reusable folders of instructions, examples, and optional scripts that teach Claude how to complete a specific task the same way every time. Anthropic calls the feature Agent Skills, commonly written as Claude Agent Skills since it is built specifically for the Claude model family, and it works by keeping most of that folder out of Claude's context window until a task actually calls for it. A Skill typically contains a SKILL.md file with a short description plus deeper how-to steps, and it can bundle executable code so Claude runs a script instead of reasoning the task out from scratch. The result is a single package of expertise that behaves the same whether it is invoked inside Claude.ai, Claude Code, or an API-based agent.
Key Stats
- Gartner projects that by 2028, 33 percent of enterprise software applications will include agentic AI capabilities, up from less than 1 percent of applications in 2024 (Gartner, 2025).
- McKinsey's global AI survey found that 65 percent of organizations reported regularly using generative AI in at least one business function (McKinsey, 2024).
What Are Claude Skills?
Claude Skills are packaged units of task-specific know-how that Claude reads only when a job requires them, instead of instructions crammed permanently into a prompt. Anthropic released the feature in October 2025 under the formal name Agent Skills, describing it as a way to give Claude real-world expertise: how a specific company formats a report, how a team wants its spreadsheets built, or how a support workflow should run step by step. This kind of packaged, reusable capability is part of the same shift Gartner points to in projecting that agentic AI features will appear in a third of enterprise software by 2028, since Skills are one concrete way vendors are building that repeatable, task-specific behavior into AI products today. Each Skill lives in its own folder and can be shared, versioned, and reused across projects much like a code library. Because the format is plain text and files, a marketing team, a legal department, and an engineering group can each maintain their own Skills library without touching the underlying model.
How Do Claude Skills Work?
Claude Skills work through a loading pattern Anthropic calls progressive disclosure, which reveals a Skill's content in layers rather than all at once. At the first layer, Claude always has access to just the name and description of every available Skill, a small enough footprint that hundreds of skills can sit on the shelf without consuming meaningful context. When a request matches what a Skill is described to do, Claude reads the full SKILL.md file for detailed instructions, and only then, if the Skill includes scripts or reference files, does it open or run those as a final layer. This staged approach is what separates Claude Skills from simply pasting a long document into a system prompt: the model spends tokens only on the expertise the current task actually needs.
What Is Inside a Claude Skill?
A Claude Skill is a folder, and at minimum it holds one file called SKILL.md, written as plain Markdown with a small header containing the Skill's name and description. Below that header sits the instruction body: step-by-step guidance, formatting rules, examples, or decision logic written in plain language. Many Skills go further and bundle supporting material in the same folder, such as templates, checklists, reference documents, or small scripts that Claude can run directly rather than working out the same logic from scratch every time. This bundling matters in practice: a Claude Code skill for building a spreadsheet, for example, can ship the actual formula generation script alongside the instructions, so output stays consistent across runs instead of drifting each time the model improvises.
How Are Claude Skills Different From a System Prompt?
A system prompt is always loaded and always active, while a Claude Skill is dormant until the task calls for it, and that is the core practical difference between the two. A system prompt has to describe every behavior a user might ever need up front, which gets bloated and expensive in tokens as an assistant takes on more responsibilities. A Skill instead sits alongside dozens of other Skills, invisible to the context window until its description matches the current request, so an organization can maintain a large library of narrow, specific behaviors without paying a constant token cost for all of them. In short, a system prompt describes who Claude is in general, while a Skill describes how to do one particular job well.
How Do Claude Skills Compare to MCP and Other Approaches?
Claude Skills and the Model Context Protocol (MCP) solve different problems and are designed to work together rather than compete. MCP, which Anthropic open-sourced in November 2024, is a protocol for connecting Claude to external systems such as databases, internal tools, SaaS products, and other live data sources it would otherwise have no access to. A Skill, by contrast, does not grant new access. It packages the know-how for using access Claude already has, including any MCP connected tools, in a specific, repeatable way. A team might use MCP to give Claude a live connection to its CRM, then use a Skill to teach Claude exactly how that team wants a weekly pipeline report formatted from the data MCP returns.
| Approach | What it provides | Setup effort | Portability | Best fit |
|---|---|---|---|---|
| Claude Skills | Packaged instructions, templates, and optional scripts loaded only when relevant | Low, write a SKILL.md file | Same behavior across Claude.ai, Claude Code, and the API | Repeatable, well-defined tasks and workflows |
| MCP (Model Context Protocol) | A live connection to external tools, databases, and data sources | Medium, run or connect to a server | Works with any MCP-compatible client, not only Claude | Real-time data access and third-party integrations |
| System prompt or custom instructions | Always on tone, persona, and general behavior rules | Low | Usually tied to one app or one deployment | Broad style and personality guidance |
| Fine-tuning | Retrained model weights based on custom examples | High, needs data, compute, and evaluation | Locked to that specific fine-tuned model version | Deep behavior change at large scale |
Where Can You Use Claude Skills?
Claude Skills work anywhere Claude itself runs, which is the main advantage Anthropic built into the design. That includes Claude.ai for everyday chat use, Claude Code for developers working from a terminal or IDE, and the Claude Developer Platform and Agent SDK for teams building custom agents on top of the API. A Skill written once for formatting a Word document, for example, behaves the same whether a marketer triggers it from Claude.ai or a developer triggers it from an automated Claude Code skill inside a build pipeline. Companies without an in-house AI team to design, test, and maintain a Skills library across all of these surfaces often bring in specialized LLM integration services to build and harden custom Skills so they behave consistently in production rather than only in a demo.
How Do You Create a Custom Claude Skill?
Creating a custom Claude Skill starts with a single Markdown file, not a development environment or special tooling. A developer writes a SKILL.md file with a clear name and description at the top, then documents the exact steps, rules, and examples Claude should follow for that task in the body below it. From there, optional folders for scripts, templates, or reference material can sit alongside it, and the whole folder goes wherever the platform expects it, for example a project's skills directory in Claude Code or a shared Skills library on the Claude Developer Platform. Anthropic and the developer community have also published skill creation helpers that walk through drafting, testing, and refining a new Skill's description so Claude reliably recognizes when to use it.
Are Claude Skills Secure to Use?
Claude Skills are only as secure as the source they come from, since a Skill can bundle executable code that runs with the same level of trust as any other software installed on a machine. Because of that, Anthropic recommends installing Skills only from sources you trust and reviewing a Skill's instructions and any bundled code before adding it to a shared environment, the same way a team would vet a new open source package. Enterprise deployments typically add their own review step on top of this, keeping an approved internal library of Skills rather than allowing arbitrary ones to be loaded by anyone on the team.
Frequently asked questions
Is a Claude Skill the same thing as a Claude Code skill?
A Claude Code skill is simply a Claude Skill used inside Claude Code, Anthropic's command line coding agent. The underlying format, a SKILL.md file plus optional resources, is identical everywhere Claude runs.
Do I need to know how to code to write a Claude Skill?
No. A basic Claude Skill is just a Markdown file with clear instructions, so anyone comfortable writing a how-to document can create one. Coding is only needed if you want to bundle an executable script for extra reliability.
Can Claude use more than one Skill at a time?
Yes. Claude can draw on multiple Skills within the same task when each one's description matches part of what the request needs, loading each one only as it becomes relevant.
Do Claude Skills replace the Model Context Protocol?
No. Skills and MCP solve different problems and are commonly used together, with MCP supplying live data or tool access and Skills supplying the instructions for how to use it well.
Where are Claude Skills stored?
A Claude Skill is stored as a folder, either personally, at the project level inside Claude Code, or in a shared organization library on the Claude Developer Platform, depending on who needs to use it.
Will having many Claude Skills installed slow down every conversation?
No. Because only a Skill's short name and description stay loaded by default, having many Skills available does not add meaningful overhead until one is actually triggered.
Updated July 2026. This overview reflects Anthropic's Agent Skills system for Claude as publicly documented at the time of writing and will be revisited as the feature evolves.
Want custom Claude Skills built and hardened for production? See Codioo's LLM integration services.