What Are Claude Skills? The Complete Guide

Claude Skills are organized folders of instructions, scripts, and resources that Claude loads on demand to perform specialized tasks. Each Skill is a folder with a SKILL.md file at its root; Claude reads that file's name and description at startup, then pulls the full instructions into context only when a task calls for them. Anthropic shipped Skills in October 2025 and published the format as an open standard that December.

The short version: a Skill is how you hand Claude a repeatable procedure (your brand's slide format, your company's data-cleaning steps, the way you like PDFs generated) once, so you stop re-explaining it every conversation. This guide covers what Skills actually are, the mechanism that makes them cheap to keep around, how they differ from MCP and prompts, where they run, and whether they are safe.

What problem do Skills actually solve?

Large models are generalists. Ask Claude to build a spreadsheet or a deck and it will do a competent job, but it does not know your conventions, and it burns context every time you paste in the same guidance. Two bad options followed: stuff all your instructions into one enormous system prompt (expensive, and most of it is irrelevant to any given task), or fine-tune a model (slow, costly, and stale the moment your process changes).

Skills are the third option. You write the procedure down as files, and Claude reads only the parts it needs, only when it needs them. Anthropic describes a Skill as something that transforms "general-purpose agents into specialized ones" by packaging domain expertise into composable resources. The onboarding-guide analogy in the docs is the right mental model: a new hire does not read the entire company wiki on day one; they reference the relevant page when a task requires it.

That "only when needed" property is the whole trick, and it has a name.

How do Claude Skills work?

Skills run on progressive disclosure: Claude loads information in stages instead of consuming context up front. Anthropic's documentation breaks this into three levels.

Token figures are Anthropic's, from the Agent Skills overview docs (as of July 2026).

At startup, Claude sees only a Skill's name and one-line description, roughly 100 tokens. That is enough for it to know the Skill exists and when it might be relevant, without paying for content it may never use. This is why you can install many Skills without a context penalty: an idle Skill costs almost nothing.

When your request matches a Skill's description, Claude reads the full SKILL.md from the filesystem (it literally runs a bash command to open the file) and brings those instructions into context (level two). If the instructions point to other files (a detailed reference, a form-filling guide) or executable scripts, Claude reads or runs those only as the task demands (level three).

The script part matters for efficiency. When Claude runs a bundled script, the script's code never enters the context window. Only its output does. Anthropic's example: running validate_form.py costs you the words "Validation passed," not the hundred lines of Python that produced them. Deterministic code you wrote once beats asking the model to regenerate equivalent logic every time, and it does not eat your token budget.

The architecture depends on Claude having a filesystem and a code-execution environment: a VM it can navigate with bash. Skills exist as directories on that machine, and Claude reads exactly the sections each task requires, the way you would flip to one chapter of a manual.

What's inside a Skill?

The only required file is SKILL.md. It has two parts: YAML frontmatter (machine-readable metadata) and a markdown body (the human-readable instructions Claude follows).

---
name: pdf-processing
description: Extract text and tables from PDFs, fill forms, merge documents. Use when the user mentions PDFs, forms, or document extraction.
---

# PDF Processing

## Quick start
Use pdfplumber to extract text from PDFs.
For advanced form filling, see FORMS.md.

Two fields are required, and Anthropic's docs are specific about their limits:

• name: max 64 characters; lowercase letters, numbers, and hyphens only; no XML tags; and it cannot contain the reserved words "anthropic" or "claude."
• description: non-empty, max 1,024 characters, no XML tags. This is the most important line in the file: it is the only thing Claude sees at startup, so it must say both what the Skill does and when to use it. A vague description means Claude never triggers the Skill; a precise one means it fires at the right moment.

Beyond SKILL.md, a Skill folder can hold anything you reference from it. A common layout:

pdf-skill/
├── SKILL.md        # required: metadata + instructions
├── FORMS.md        # optional: extra guidance, loaded when needed
├── REFERENCE.md    # optional: detailed API reference
└── scripts/
    └── fill_form.py # optional: code Claude runs via bash

Anthropic's own repository standardizes on scripts/ for executable code, references/ for documentation, and assets/ for templates and other resources. None of it is mandatory. Many effective Skills are a single SKILL.md with no scripts at all.

Where can you use Claude Skills?

Skills work across Anthropic's products, but the surfaces differ in what they support and how you add Skills. This table reflects Anthropic's documentation as of July 2026.

A few practical wrinkles the docs flag:

• Skills do not sync across surfaces. A Skill you upload to claude.ai is not available in the API, and vice versa; Claude Code Skills are separate files entirely. You manage each surface on its own.
• Sharing scope varies. On claude.ai, custom Skills are per-user: each teammate uploads their own, and admins cannot centrally distribute them. On the API, uploaded Skills are workspace-wide. In Claude Code, Skills are personal or per-project and can also travel via Claude Code Plugins.
• Network access varies. API Skills run with no internet access and no runtime package installs. Claude Code Skills have the same network access as any program on your machine. On claude.ai it depends on your admin's settings.

If you spend your time in a terminal, Claude Code is the easiest entry point: a Skill is just a folder, no upload step. That workflow sits alongside the coding tools we cover in Cursor vs Claude Code and the broader best AI coding tools roundup.

Claude Skills vs MCP vs prompts vs subagents

Skills get conflated with three other Claude features. They solve different problems, and Anthropic frames the distinctions cleanly.

The Skills-versus-MCP question is the one people actually ask. The clean line: MCP gets Claude to the data; a Skill tells Claude what to do once it is there. MCP is plumbing: it connects Claude to your CRM. A Skill is the runbook: it teaches Claude how your team turns that CRM data into a monthly report. They compose. Use MCP to reach the database, then a Skill to standardize the analysis and the output format.

Prompts and Skills differ on repetition. A prompt is guidance for one conversation; a Skill is guidance that loads automatically across every conversation where it is relevant, so you write it once. If you find yourself pasting the same instructions into Claude for the third time, that is the signal to turn them into a Skill.

What pre-built Skills ship today?

Anthropic maintains a set of pre-built Skills that Claude uses automatically: no setup, no authoring. The launch set covers office documents:

• PowerPoint (pptx): create and edit slides, analyze deck content
• Excel (xlsx): build spreadsheets with formulas, analyze data, generate charts
• Word (docx): create and format documents
• PDF (pdf): generate formatted PDFs and fill forms

These are why, when you ask claude.ai to "make me a deck," a real .pptx comes back: a Skill is doing the work behind the scenes. As of 2026 these same document Skills are surfacing inside Anthropic's Claude for Excel and PowerPoint add-ins for Microsoft 365, which let Claude read and modify spreadsheets and slides in place, based on Anthropic's product documentation.

Anthropic also publishes open-source example Skills in its public skills repository, spanning creative, technical, and enterprise categories, plus a template for authoring your own. Note the licensing split the repo calls out: the document-creation Skills (docx, pdf, pptx, xlsx) are source-available rather than open source, while many of the other example Skills are Apache 2.0.

Skills are now an open standard

In December 2025, Anthropic published Agent Skills as an open standard and released the specification for any AI platform to adopt; it is governed as a cross-platform spec at agentskills.io. The practical consequence: SKILL.md is no longer a Claude-only format. A Skill you write is portable to other agent tools that implement the standard, and a growing library of partner-built Skills (from companies like Atlassian, Figma, Canva, Notion, and others) targets the same format.

This matters strategically more than it looks. Formats that become standards tend to win the ecosystem, the same pattern Anthropic ran with MCP a year earlier. If you are investing time in authoring Skills, you are writing to a spec with adoption beyond a single vendor, not a proprietary config file that dies if one product does.

Are Claude Skills safe to use?

Only when you trust the source. This is the one part of Skills you cannot treat casually. A Skill gives Claude new instructions and code, and, in Anthropic's own words, "a malicious Skill can direct Claude to invoke tools or execute code in ways that don't match the Skill's stated purpose." Depending on what Claude can access when the Skill runs, the downside includes data exfiltration and unauthorized system access.

Anthropic's guidance, paraphrased:

• Treat installing a Skill like installing software. Use Skills you wrote yourself or got from Anthropic. Anything from an unknown source gets audited before it runs.
• Read every file in the bundle (SKILL.md, scripts, and any resources) for patterns that do not match the stated purpose: unexpected network calls, odd file access, tool use that has nothing to do with the job.
• Watch external fetches. Skills that pull content from external URLs are the highest-risk category, because fetched content can itself carry malicious instructions, and even a trustworthy Skill can be compromised if its dependencies change later.

The risk is not hypothetical hand-waving; it is the standard supply-chain problem applied to agent instructions. The mitigation is boring and effective: know where your Skills come from.

Who should use Claude Skills, and for what?

Skills earn their keep whenever a task is (a) repeatable and (b) has a right way to do it that Claude does not know by default. Some concrete fits:

• Standardized documents. Brand-compliant decks, reports in your house format, contracts from your template. A Skill encodes the format once.
• Data workflows. Your specific cleaning steps, validation rules, and reporting layout, so every analysis comes out consistent.
• Content and SEO operations. A Skill can carry your editorial rules, internal-linking conventions, or schema requirements, the kind of repeatable production work behind programmatic SEO and the on-page discipline in our LLM SEO guide. Encoding your SEO checklist as a Skill means Claude applies it the same way every time instead of you re-explaining it per article.
• Engineering chores. Test scaffolding, migration patterns, deploy runbooks: the procedural knowledge a senior engineer carries in their head. If you are already building with AI, the workflow pairs naturally with the approach in our vibe coding tutorial.

Skills are less useful for genuinely one-off tasks (just prompt), for connecting to external systems (that is MCP's job), or for anything where the "right way" changes every time and cannot be written down.

What to do next

If you want to try one without authoring anything, open claude.ai and ask it to build a spreadsheet or a slide deck. You are already using the pre-built Skills. To go a level deeper, the fastest sandbox is Claude Code, where a Skill is just a SKILL.md file in a folder and nothing has to be uploaded.

The core idea is small and worth internalizing: a Skill is procedural knowledge, stored as files, loaded only when relevant. Once that clicks, the question stops being "what are Claude Skills" and becomes "which of my repeated instructions should be one." For the broader toolkit around building with Claude, start with best AI coding tools.