What Are Claude Skills? Agent Skills Explained

Claude Skills — officially called Agent Skills — are folders of instructions, scripts and reference files that Claude loads on its own, mid-conversation, when a task calls for them. Instead of re-explaining a workflow every time, you (or Anthropic) package it once as a SKILL.md file plus any supporting code, and Claude reads it from the filesystem only when it is relevant, then runs it like a specialist following a playbook. Skills work the same way across claude.ai, Claude Code and the Claude API, and Anthropic ships four built-in ones for PowerPoint, Excel, Word and PDF files.

What Are Claude Skills?

A Skill is, structurally, nothing exotic: it's a directory that contains a required SKILL.md file and, optionally, extra markdown references and executable scripts. Anthropic's engineering team describes the idea as packaging expertise into a composable resource — the kind of onboarding document you'd hand a new hire, except Claude can read it instantly and follow it exactly. A Skill might teach Claude how to fill out a specific PDF form, how to apply a company's brand guidelines to a slide deck, how to structure meeting notes in a particular format, or how to file a ticket in Jira using your team's conventions.

What makes Skills different from just pasting instructions into a chat is that they persist and travel. Build a Skill once and it's available in every future conversation where it's installed, and — because the format is the same everywhere — the same SKILL.md can run in claude.ai, Claude Code or through the API without rewriting it. Anthropic positions this against the alternative of constantly re-explaining the same workflow, or stuffing a system prompt with every possible instruction "just in case," which wastes context on guidance the model may never need.

How Claude Finds and Loads a Skill

The mechanism behind Skills is what Anthropic calls progressive disclosure, and it's the reason you can install dozens of Skills without bloating every conversation. Skills run inside Claude's code-execution environment, where the model has real filesystem access and can use bash to read files on demand — so a Skill doesn't have to be loaded all at once. It loads in three stages:

• Level 1 — Metadata (always loaded): just the name and description fields from each Skill's YAML frontmatter go into the system prompt at startup, at roughly 100 tokens per Skill. This is what Claude scans to decide whether a Skill might be relevant to your request.
• Level 2 — Instructions (loaded when triggered): if your request matches a Skill's description, Claude runs a bash command to read the full body of SKILL.md — typically under 5,000 tokens of procedural guidance — into context.
• Level 3 — Resources and code (loaded as needed): a Skill can bundle extra reference files (an API spec, a database schema, a style guide) and scripts. Claude only reads a reference file if the task actually needs it, and when it runs a script, only the script's output — not its source code — enters the context window.

The practical effect, per Anthropic's documentation, is that a Skill can ship comprehensive documentation or large datasets with no context penalty for the parts that go unused. A PDF-processing Skill, for instance, might include a short quick-start guide plus a much longer FORMS.md reference — Claude only opens the forms file if the task actually involves filling out a form.

The Built-In Skills: PowerPoint, Excel, Word and PDF

Anthropic pre-built four Agent Skills for the document types people ask about most often, identified by short IDs: pptx, xlsx, docx and pdf. They cover creating presentations and editing slides, building spreadsheets and generating charts and reports, drafting and formatting Word documents, and producing or manipulating PDFs — including tasks like extracting form fields without loading an entire document into context. These are active by default whenever you ask claude.ai, or an API integration with the code execution tool enabled, to create or edit one of those file types; no setup is required on the user's end. The underlying code for the document Skills is source-available in the anthropics/skills repository, though — unlike the example Skills, which are Apache 2.0 — it isn't open source.

For a side-by-side look at how Microsoft's own AI assistant handles the same Office file types, see Pandromeda's explainer on ChatGPT in Microsoft Word, Excel and PowerPoint.

Claude Skills vs. a Plain Prompt

A prompt — or a saved custom instruction — is conversation-level: you type it, it shapes that one exchange (or, for a saved system prompt, every exchange), and the model has to re-read and weigh it every single time regardless of whether it's relevant. A Skill is different in two ways. First, it's conditional: it only enters the context window when Claude's own matching against the Skill's description decides the current task actually needs it, so you can have many Skills installed with almost no ongoing token cost. Second, it can carry executable code and reference material a prompt cannot — a Skill can bundle a Python script that Claude runs deterministically, rather than having the model re-generate equivalent logic from scratch on every request, which Anthropic notes is both more reliable and cheaper in tokens. In short: a prompt tells Claude what to do once; a Skill teaches Claude a repeatable procedure it can recognize and reuse on its own.

Claude Skills vs. AI Agents and MCP Tools

Skills, agents and tools solve different problems and are meant to be combined rather than compared as substitutes. An AI agent is the broader pattern: a model that plans, takes actions, calls tools and iterates toward a goal with some autonomy — Pandromeda has covered that concept in detail in What Is an AI Agent, and How Does It Actually Work? Tools (including those exposed through the Model Context Protocol, or MCP) are what give an agent the ability to act — send an email, query a database, hit an API. A Skill doesn't give Claude a new capability like a tool does; it gives Claude know-how about how to use the capabilities it already has for a specific, recurring job. Anthropic's own framing is that Skills are composable: you can stack several of them to handle a complex, multi-step workflow, and a single agent can call on multiple Skills plus multiple tools in the same task. A useful rule of thumb: tools are the hands, Skills are the expertise, and an agent is the whole worker using both.

How to Create a Custom Skill

Every Skill needs, at minimum, a SKILL.md file with YAML frontmatter containing two required fields:

• name — up to 64 characters, lowercase letters, numbers and hyphens only; it cannot contain XML tags or the words "anthropic" or "claude."
• description — up to 1,024 characters, and the single most important field to get right, since it's what Claude matches your request against to decide whether to trigger the Skill. Anthropic's authoring guidance says it should state both what the Skill does and when to use it.

Below the frontmatter, the body of SKILL.md is plain markdown instructions — steps, examples, guidelines — and you can reference additional files (a longer REFERENCE.md, a scripts/ folder) that Claude will open only when the task calls for them. Anthropic's suggested workflow for building one is iterative: find the gaps in Claude's performance on a representative task, write up the missing procedure, split it into multiple files once it grows complex, and — usefully — you can collaborate with Claude itself to capture a successful one-off solution into a reusable Skill.

Where you create it depends on the surface. In Claude Code, a custom Skill is just a directory dropped into ~/.claude/skills/ (personal, available in every project) or a project's .claude/skills/ folder — no upload step, Claude discovers it from the filesystem. In claude.ai, you upload a Skill as a zip file under Settings > Features, and it applies only to your own account (there's no org-wide admin distribution on claude.ai). Through the API, you upload a Skill — as a zip or individual files — to the /v1/skills endpoint, which returns a skill_* ID; API Skills are shared workspace-wide rather than per-user. Anthropic is explicit that custom Skills do not sync across these three surfaces: a Skill uploaded to claude.ai has to be uploaded separately to the API, and vice versa.

Where Claude Skills Are Available

Pre-built document Skills and custom Skills aren't available in identical form everywhere. The table below summarizes how each surface differs, based on Anthropic's Agent Skills documentation.

One detail worth flagging for anyone building production workflows: Agent Skills is not covered by Anthropic's zero data retention (ZDR) arrangements, so Skill definitions and execution data follow the company's standard retention policy rather than a ZDR exception. For current model pricing and access tiers that interact with Skills on claude.ai, see Pandromeda's coverage of Claude Opus 5.5 pricing and access, and for how Claude Code's own subscription compares with rival coding tools, see Cursor vs. GitHub Copilot vs. Claude Code vs. Codex.

Is There a Claude Skills Marketplace?

Not an official storefront in the App Store sense, but Anthropic does publish a public catalog of example Skills on GitHub at anthropics/skills, organized into categories — creative and design, development and technical, enterprise and communication, plus the document Skills themselves — alongside a skill-creator template and the Agent Skills specification. In Claude Code specifically, that repository can be registered as a plugin marketplace with /plugin marketplace add anthropics/skills, after which individual skill bundles install with a command like /plugin install document-skills@anthropic-agent-skills. Anthropic has also said Skills were released as an open specification in December 2025, which is what has let a wider ecosystem of third-party Skill directories and community repositories spring up since launch — those are not Anthropic products, so the same trust rules below apply to them.

Security: Treat Skills Like Installing Software

Because a Skill can bundle executable code and instructions that steer Claude's use of tools like bash or file access, Anthropic's guidance is blunt: only use Skills from trusted sources, meaning ones you wrote yourself or ones Anthropic published. A malicious or compromised Skill could direct Claude to exfiltrate data, access systems it shouldn't, or misuse tools in ways that don't match what the Skill claims to do — and Skills that fetch content from external URLs carry extra risk, since that fetched content could itself smuggle in instructions. Anthropic's recommendation for anyone who must use a Skill from an unfamiliar source is to audit every bundled file — the SKILL.md, any scripts, any images — for network calls or file operations that don't match the stated purpose, and to treat installing a Skill with the same caution as installing any other third-party software, especially in a production system with access to sensitive data. Claude Enterprise customers can additionally turn on Skill content scanning for custom Skills uploaded in claude.ai and Claude Cowork, though that scanning doesn't extend to Skills uploaded via the API or Console.

What's Next

Agent Skills launched as a product feature in October 2025 and became an open specification two months later, which suggests Anthropic wants Skills treated less like a proprietary claude.ai feature and more like a portable format other tools and vendors can adopt — not unlike how MCP spread beyond Anthropic's own products after its release. Expect the near-term developments to center on closing the gaps that exist today: Skills still don't sync automatically between claude.ai, Claude Code and the API, there's no centralized admin console for distributing them across a claude.ai team, and the built-in document Skills remain absent from Claude Code. Anthropic's own enterprise documentation on governance and content scanning points to where the next round of controls is likely to land, for organizations that want to vet and deploy Skills at scale rather than relying on individual users to audit each one by hand.

Written byTheo Park

Theo Park runs the AI desk at Pandromeda. He follows model launches from the frontier labs and the open-weight community, tracks the assistants and developer tools built on them, and explains what each release changes on pricing, capability and safety. His reporting leans on primary sources: model cards, technical reports, API documentation and the companies' own announcements.

More from Theo ParkAll AI coverage