Citation and evidence

Claude Skills

46 min full readUpdated 38 references

This article's verification

Report a problem with this article

Edit article
More

Use this article

Raw MarkdownExplore connections

Improve this page

Revision historyDiscuss this article

Topics

AI Tools & ProductsAnthropic

Cite this article

Claude Skills (also marketed as Agent Skills) are a feature developed by Anthropic that lets users package repeatable workflows, instructions, reference materials, and optional executable scripts into discoverable folders that Claude can load on demand. Each Skill is a directory on a filesystem that contains a SKILL.md file with YAML frontmatter, plus any supporting files the workflow needs. Anthropic announced Skills publicly on October 16, 2025 and released them at the same time across Claude.ai, Claude Code, the Claude API, and the Claude Agent SDK.[1][2]

The central design idea is progressive disclosure, which Anthropic calls "the core design principle that makes Agent Skills flexible and scalable."[3] Claude is told only the name and one-line description of every installed Skill at the start of a session, then reads the full SKILL.md only when the user's request matches, and only opens additional bundled files (templates, references, scripts) when a step actually needs them. This keeps the context window small while letting a single account install many specialized capabilities.[3][4] Independent developer and AI commentator Simon Willison wrote on the day of launch that Skills "might be a bigger deal than MCP," pointing to the simplicity of the format compared with the Model Context Protocol.[5] Two months later, on December 18, 2025, Anthropic released the format as an independent open standard at agentskills.io, opened organization-wide admin provisioning to Team and Enterprise plans, and launched a curated partner-skills directory anchored by Atlassian, Canva, Cloudflare, Figma, Notion, Sentry, Vercel, and Zapier.[20][21][23] On March 31, 2026 that directory became a unified surface at claude.ai/customize covering Skills, connectors, and plugins.[24] GitHub Copilot, VS Code, and Cursor already supported the format at the open-standard launch and OpenAI Codex added it within days; by November 2026 the agentskills.io client showcase listed more than forty products, including Gemini CLI.[20][21][22] On October 22, 2026, Anthropic donated Agent Skills to the Linux Foundation's Agentic AI Foundation.[36]

Quick reference

AttributeValue
Developed byAnthropic
Initial releaseOctober 16, 2025
Open standard releaseDecember 18, 2025
Standard homeagentskills.io (an Agentic AI Foundation project since October 22, 2026)
Unified directoryclaude.ai/customize (skills, connectors, and plugins)
File formatFolder containing SKILL.md with YAML frontmatter plus optional resources
Required fieldsname, description
Available onClaude.ai Free/Pro/Max/Ultra/Team/Enterprise (as of November 2026), Claude Code, Claude API, Claude Agent SDK, Claude Cowork
Reference repositorygithub.com/anthropics/skills
License (most skills)Apache 2.0
License (document skills)Source-available, not open source
Notable adoptersGitHub Copilot, VS Code, Cursor, Gemini CLI, OpenAI Codex, Goose, OpenCode, Amp, Letta, Junie, Kiro, Roo Code

Expanded article table

"Agent Skills: Specialized capabilities you can customize", official video from Anthropic, published October 16, 2025. Watch on YouTube.

What is a Claude Skill?

A Skill is a folder. The minimum contents are a single SKILL.md file whose first lines are YAML frontmatter:

---
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.

## Form filling

See FORMS.md for the form-filling reference.

The name field is the unique identifier (lowercase, hyphenated, maximum 64 characters, and not allowed to contain the words "anthropic" or "claude"). The description is a short string of at most 1,024 characters that tells Claude both what the Skill does and when to use it. Anything below the closing --- is Markdown that Claude reads as instructions when the Skill is triggered.[2][6]

A more complex Skill folder may look like this:

pdf-skill/
├── SKILL.md              # main instructions
├── FORMS.md              # extra reference loaded only when needed
├── REFERENCE.md          # detailed API reference
├── scripts/
│   └── fill_form.py      # executable Python script
└── assets/
    └── template.pdf      # binary template

Claude reaches into these files using ordinary bash commands inside the code-execution sandbox: it runs cat SKILL.md to load the main file, then cat FORMS.md if a step references it, then executes python scripts/fill_form.py for deterministic operations. Script source code never enters the context window; only the printed output of the script does, which makes Skills far more token-efficient than asking Claude to write the same logic from scratch on every call.[3][4]

Anthropic likens the architecture to an onboarding binder for a new employee, writing that "building a skill for an agent is like putting together an onboarding guide for a new hire."[3] The table of contents (Skill metadata) is always visible, individual chapters (the SKILL.md body) get pulled off the shelf only when a question lands in their territory, and the appendices (scripts, templates, reference files) only get opened when a specific step demands them. Skills are not always-on personality settings; they are situation-triggered reference material that costs almost nothing to keep around.

How does the Skills architecture work?

How does progressive disclosure work?

Progressive disclosure is the core engineering pattern. The official documentation describes three loading levels:[3][6]

LevelContentWhen loadedApproximate token cost
1. Metadataname and description from YAML frontmatter of every installed SkillAt session startup, included in the system promptAbout 100 tokens per Skill
2. InstructionsFull body of SKILL.md for one SkillOnly when Claude judges the Skill is relevant to the current requestUnder about 5,000 tokens per Skill
3. ResourcesAdditional Markdown files, scripts, templates, datasetsOnly when a step references themEffectively unlimited; bundled files cost zero tokens until accessed

Expanded article table

The consequence is that a user can install dozens of Skills without bloating the context window. Claude knows each Skill exists and what it is for, but it does not pay the token cost of any Skill it does not actually use. When the model finally does call a Skill, it pulls only the specific files that step needs.

A representative load sequence: at startup the system prompt includes the line pdf-processing - Extract text and tables from PDF files. The user types "Extract the text from this PDF and summarize it." Claude matches the request to the description, issues bash: cat ~/.claude/skills/pdf-processing/SKILL.md, and the body enters context. Claude reads the body, notices that form-filling is not needed, skips FORMS.md, then follows the instructions and calls any bundled scripts via bash.[6]

Claude does not re-read the SKILL.md file on each turn after activation. The rendered content sticks in the conversation as a single message, and Claude Code carries the most recent invocation forward across auto-compaction within a 25,000-token shared budget for re-attached Skills.[15]

Where do Skills live on disk?

Skills live as ordinary directories on a virtual machine, accessible to Claude over the same bash interface it uses for any other code-execution task. Because Skills are just folders, version control is straightforward: a Skill can be a Git repository, distributed as a ZIP file, or committed to a project's .claude/skills/ directory. Any tool that can read Markdown and run shell commands can in principle execute a Skill.[2][7] This decouples Skills from the model: the same folder works whether Claude is a Pro user on claude.ai, an API caller running batch jobs, or a developer in Claude Code iterating in a terminal.

What fields go in the YAML frontmatter?

The frontmatter is the only part of a Skill that Claude is forced to read. Anthropic enforces several constraints on its fields, and Claude Code adds a number of optional fields specific to its runtime:[6][15]

FieldRequiredSurfaceRules
nameYes (API), Optional (Claude Code)AllLowercase letters, digits, and hyphens only. Maximum 64 characters. May not contain XML tags or the reserved words "anthropic" and "claude". Defaults to the directory name in Claude Code if omitted.
descriptionYes (recommended)AllNon-empty, maximum 1,024 characters, no XML tags. Describes what the Skill does and when Claude should use it.
versionNoNot a standard fieldNot part of the Agent Skills spec; the spec suggests recording a version inside metadata. API pinning uses versions assigned by the Skills API (date-based for Anthropic Skills, version IDs for custom Skills), not frontmatter.[27][37]
when_to_useNoClaude CodeAdditional trigger phrases or example requests appended to the description in the Skill listing.
argument-hintNoClaude CodeHint shown during autocomplete (for example [issue-number]).
argumentsNoClaude CodeNamed positional arguments for $name substitution in the body.
allowed-toolsNoClaude Code, open standard (experimental)Tools Claude can use without prompting permission while the Skill is active. The open standard exposes the same field as an experimental capability gate.[27]
disable-model-invocationNoClaude CodeIf true, only the user can invoke with /<name>; Claude does not auto-trigger.
user-invocableNoClaude CodeIf false, hidden from the slash command menu.
modelNoClaude CodeModel override for the rest of the current turn.
effortNoClaude CodeEffort level override (low, medium, high, xhigh, max).
contextNoClaude CodeSet to fork to run in a forked subagent.
agentNoClaude CodeSubagent type to use when context: fork is set.
pathsNoClaude CodeGlob patterns that gate auto-activation by file path.
hooksNoClaude CodeHooks scoped to this Skill's lifecycle.
shellNoClaude Codebash (default) or powershell for inline command injection.
licenseNoOpen standardLicense name or a reference to a bundled license file (for example Apache-2.0). Records the Skill's redistribution terms in a portable way across implementations.[27]
compatibilityNoOpen standardMaximum 500 characters of free text describing environment requirements (intended product, system packages, network access), for example "Designed for Claude Code (or similar products)".[27]
metadataNoOpen standardArbitrary string-to-string map for vendor-specific extensions. Authors are asked to namespace keys to avoid collisions.[27]

Expanded article table

Writing a precise description is the single most important authoring task. Because Claude only sees the description when deciding whether to load a Skill, a vague line like "helps with documents" will produce inconsistent activation, while a specific line like "Extract text and tables from PDF files; use when the user mentions PDFs, forms, or document extraction" will trigger reliably.[3][6] If a description is unhelpful, Anthropic recommends putting the most distinctive use case first because the combined description and when_to_use text is truncated at 1,536 characters in the Claude Code listing to limit context usage.[15]

How do discovery scopes resolve conflicts?

Claude Code loads Skills from several locations; four of them have a defined precedence or namespace when the same name appears in more than one place:[15]

ScopePathApplies toPrecedence
Enterprise.claude/skills/<name>/SKILL.md in the managed settings directoryAll users in the organizationHighest
Personal~/.claude/skills/<name>/SKILL.mdAll projects on a user's machineSecond
Project.claude/skills/<name>/SKILL.md (in repo root)One repository, all collaboratorsThird
Plugin<plugin>/skills/<name>/SKILL.mdWherever the plugin is enabledNamespaced as plugin-name:skill-name, no conflicts

Expanded article table

The API uses a different model: custom Skills uploaded via the /v1/skills endpoint are workspace-wide and visible to every member of the workspace, while Claude.ai uploads start private to the uploading user; on Team and Enterprise plans users can share them with colleagues or publish them to the organization, and owners can provision Skills for everyone.[24][35] Claude Code also auto-discovers Skills from nested .claude/skills/ directories within a project, which supports monorepos where individual packages carry their own Skills, and from any directory passed via the --add-dir flag at startup.[15]

When a name collision happens (for example a project Skill called commit and a personal Skill called commit), enterprise overrides personal and personal overrides project. Plugin Skills sit in a namespace of their own and are addressed as plugin-name:skill-name, so they cannot collide with anything else.

Live change detection

Claude Code watches Skill directories for filesystem changes during a session. Adding, editing, or removing a Skill under any watched path takes effect within the current session without a restart. The exception is creating a top-level Skills directory that did not exist when the session started: Claude Code does not watch that directory, so the /reload-skills command is needed to pick up Skills placed there.[15] The hot-reload behavior shipped in Claude Code 2.1.0 on January 7, 2026, alongside the /teleport command and a response-language setting.[38]

How does Claude load a Skill?

The lifecycle of a Skill within a single Claude conversation looks like this:[3][4][6]

  1. Startup. Claude's system prompt is augmented with the name and description of every installed Skill. The full SKILL.md files are not loaded.
  2. User request. Claude examines the request against the descriptions of installed Skills.
  3. Skill trigger. If a description matches, Claude issues a bash call such as cat ~/.claude/skills/pdf-processing/SKILL.md. The body enters the context window.
  4. Reference files. If SKILL.md mentions another bundled file, Claude reads it with another bash call only when that branch of the workflow is needed.
  5. Script execution. When SKILL.md instructs Claude to run a script, Claude executes it via bash. Only the script's stdout/stderr enters context, not the source code.
  6. Composition. A single user request can trigger multiple Skills. Anthropic emphasizes that Skills are composable: Claude can stack a brand-guidelines Skill on top of a pptx Skill to produce a branded presentation, with no manual orchestration required.[1][2]

Because every step is just bash against ordinary files, there is no proprietary protocol layer between the model and the Skill. Claude treats Skills the way a developer would treat a folder of helper scripts. That is what makes Skills feel different from a Custom GPT or a Gemini Gem: no hosting layer to depend on, no upload step to gate, just files that the model can read on demand.

Dynamic context injection

Claude Code Skills support a preprocessing feature called dynamic context injection. The syntax !`<command>` inside SKILL.md is replaced by the command's output before Claude sees the file. A Skill that summarizes uncommitted changes might use:

## Current changes
!`git diff HEAD`

## Instructions
Summarize the changes in two or three bullet points.

When Claude loads this Skill, the live diff has already been substituted in. The model receives the actual changed lines, not the literal command string. Multi-line variants use a fenced code block opened with three backticks plus an exclamation mark. The pattern is preprocessing, not tool use, so it consumes the cost of the substitution but never enters the agent loop.[15]

What are Claude Skills used for?

Anthropic and early adopters have published a range of representative use cases:[1][2][3]

Use caseExampleWhat the Skill encodes
Brand voiceA brand-guidelines Skill applied to every generated documentApproved colors, typography, tone-of-voice rules, banned phrasing
Internal documentationA company-onboarding Skill that answers new-hire questionsOrg chart, process diagrams, policy summaries, internal links
Output formatsAn xlsx Skill that produces consistent Excel reportsCell formatting templates, chart conventions, formula libraries
Reporting workflowsA quarterly-report Skill for finance teamsStep-by-step procedure for combining spreadsheets, anomaly checks, narrative templates
Design assetsA slack-gif-creator Skill that produces optimized animated GIFsSize limits, palette constraints, FFmpeg command snippets
EngineeringA webapp-testing Skill that runs Playwright against a target siteTest scaffolding, common selectors, screenshot capture commands
Developer toolingAn mcp-builder Skill that scaffolds new MCP serversProject layout, registration boilerplate, TypeScript or Python templates
Personal automationA recipe-formatter Skill for a home cookPreferred unit conversions, allergen tags, output template
Code reviewA commit Skill that stages and writes a conventional messageLinting checks, message format, scope rules
ComplianceA legal-review Skill applied to contract draftingClause libraries, banned language, jurisdiction notes

Expanded article table

The broader pattern is that any workflow a knowledge worker repeats becomes a candidate for a Skill: the natural-language steps become the SKILL.md body, the reference data becomes a bundled file, and any deterministic operation becomes a script. The model never had to learn the workflow during training. It just reads the manual you wrote, the same way a contractor on day one would.

What Skills ship built in?

Anthropic ships a set of pre-built Skills directly from the Claude API and Claude.ai. These are sometimes called document Skills because they target the four common Office document formats:[6][8]

skill_idFormatCapabilities
pptxPowerPointCreate presentations, edit slides, apply layouts and templates, analyze existing decks
xlsxExcelCreate spreadsheets, write formulas, generate charts, perform data analysis
docxWordCreate documents, edit content, preserve formatting, work with tracked changes and comments
pdfPDFGenerate formatted PDFs, extract text and tables, fill forms, merge and split files

Expanded article table

As of November 2026 these Skills are available to all Claude.ai users, including the Free plan, and to API callers who include the code execution tool.[35][37] They activate automatically when a user asks Claude to make or modify the corresponding file type.[6] Crucially, document Skills are licensed source-available rather than open source: they ship in the public anthropics/skills repository so that users can read them, but Anthropic reserves the right to limit redistribution and derivative works, unlike the rest of the repository which is Apache 2.0.[7]

Anthropic also bundles a claude-api Skill that gives Claude up-to-date reference material for the Claude API and SDKs across eight programming languages, and ships a skill-creator Skill that walks users through authoring new Skills interactively.[6][7] Claude Code ships with several additional bundled Skills, most available in every session, such as /doctor, /code-review, /batch, /debug, /loop, and /claude-api. Unlike most built-in commands, which execute fixed logic, these are prompt-based: they hand Claude detailed instructions and let it orchestrate the work using its tools.[15]

What is in the Anthropic skills repository?

Anthropic publishes the open-source anthropics/skills GitHub repository as a public reference catalog and as an installable Claude Code marketplace.[7] The repository is registered as a Claude Code plugin marketplace using the command:

/plugin marketplace add anthropics/skills

From there, users can install bundled Skill packs:

/plugin install document-skills@anthropic-agent-skills
/plugin install example-skills@anthropic-agent-skills

As of mid-2026 the repository shipped seventeen Skills, organized loosely into four categories (by November 2026 it had nineteen, after academy-guide and discernment-nudge were added):[7]

SkillCategoryDescription
algorithmic-artCreativeGenerate p5.js art with seeded randomness and reproducible parameters
canvas-designCreativeProduce PNG and PDF designs from natural-language descriptions
slack-gif-creatorCreativeBuild optimized animated GIFs sized for Slack reactions
theme-factoryCreativeGenerate cohesive themes for Claude Artifacts and other UI
frontend-designCreativeProduce React/Tailwind component sketches and mockups
web-artifacts-builderDevelopmentScaffold complex Claude.ai HTML Artifacts using React, Tailwind, and shadcn/ui
mcp-builderDevelopmentScaffold new Model Context Protocol servers
webapp-testingDevelopmentRun Playwright test scaffolds against a target site
claude-apiDevelopmentProvide up-to-date Claude API and SDK reference material across eight programming languages
brand-guidelinesEnterpriseApply organizational color, typography, and tone-of-voice rules
internal-commsEnterpriseDraft consistent internal newsletters and announcements
doc-coauthoringEnterpriseCollaborative document drafting with revision tracking
pdfDocumentPDF extraction, generation, and form filling
docxDocumentWord document creation and editing with formatting fidelity
pptxDocumentPowerPoint presentation creation and editing
xlsxDocumentExcel spreadsheet creation, formulas, charting
skill-creatorMetaInteractive Skill authoring assistant

Expanded article table

The repository also publishes the template/ directory (a minimal scaffolding starter for new Skills) and the spec/ directory containing the Agent Skills standard specification.

Community catalogs

A community ecosystem grew quickly. Catalogs such as awesome-claude-skills, claude-skills-marketplace, tonsofskills.com, and SkillsMP collect community Skills, and several open marketplaces aggregated thousands within months.[9][10][11] Vercel's skills.sh directory, launched in early 2026, lists tens of thousands. Doppio.ai's Claude Code Skills Directory and the VoltAgent awesome-agent-skills repository (a cross-vendor list covering Claude Code, Codex, Gemini CLI, and Cursor) round out the most-cited curated indexes, with the VoltAgent list alone curating more than 1,000 Skills.[29] Because Skills are plain folders, they can be installed by cloning a repository, copying a ZIP into a settings menu, or committing a .claude/skills/ directory to a project. There is no central approval process and no revenue-sharing storefront comparable to the GPT Store, which means the community ecosystem is messier but also lower-friction.

The largest single community effort is Superpowers, a meta-framework by Jesse Vincent ("obra") that bundles a library of composable Skills (fifteen as of November 2026, including brainstorming, writing-plans, and executing-plans) with a bootstrap loader.[30] Released in October 2025 and added to Anthropic's official plugin marketplace by January 2026, Superpowers chains Skills across the software lifecycle (brainstorming, git-worktree setup, plan writing, subagent execution, TDD, and code review). It had about 312,000 stars and 27,800 forks by November 2026, the highest-starred Claude Skill project on GitHub, and runs unchanged in Codex CLI, Cursor, Gemini CLI, and the GitHub Copilot CLI.[30]

Skills Directory and organization deployment

Alongside the December 18, 2025 open-standard release, Anthropic launched a curated partner-skills directory and opened admin-controlled organization-wide provisioning on Team and Enterprise plans.[23] The directory let administrators push vetted Skills to every user with a single toggle. The launch lineup included Atlassian (Jira and Confluence), Canva, Cloudflare (deploying agents and MCP servers), Figma, Notion, Sentry, Vercel, and Zapier (paired with Zapier MCP, which connects to thousands of apps).[23]

On March 31, 2026 Anthropic merged the partner-skills directory, the connector list, and the plugin marketplace into a single browse surface at claude.ai/customize. The directory offers a typed filter (Skills, connectors, plugins), a search bar, and an "install" button. Users add Skills from Customize > Skills; on Team and Enterprise plans the directory also lists Skills that owners have provisioned centrally (from Organization settings > Plugins & skills) and Skills that colleagues have published to the organization.[24]

SurfaceWhat it groupsWho can installLaunched
Unified directory (claude.ai/customize)Skills, connectors, pluginsAny logged-in userMarch 31, 2026
Partner-skills sectionVetted partner-built SkillsWorkspace admins (Team, Enterprise)December 18, 2025
Plugin marketplaceMulti-component pluginsWorkspace admins; project-level installs in Claude CodeFebruary 2026
Personal uploadCustom ZIP-packaged SkillsAny user (Free plan included as of November 2026)October 16, 2025

Expanded article table

Is the Skills format an open standard?

On December 18, 2025, Anthropic released the Agent Skills format as an independent open standard hosted at agentskills.io.[20][21] The specification is deliberately small. According to Simon Willison's analysis on December 19, 2025, the entire spec is short enough to read in minutes and includes a lightly defined metadata extension field plus an experimental allowed-tools field for pre-approving the tools a Skill may use.[21][27] The published specification documents name, description, license, compatibility, allowed-tools, and metadata as the recognised frontmatter keys, directs extra properties into metadata (implementations such as Claude Code add their own fields on top), and pins the maximum description length at 1,024 characters.[27]

Microsoft had already adopted Skills in VS Code and GitHub by the December 18 launch, and OpenAI, which had quietly added the same structure to ChatGPT and Codex CLI earlier that month, added Skills to the Codex documentation on December 20.[20][21] By mid-2026 more than thirty independent agent products read the same SKILL.md files, including:[22]

VendorProduct
AnthropicClaude.ai, Claude Code, Claude Cowork
OpenAICodex CLI, ChatGPT
MicrosoftVS Code, GitHub Copilot
GoogleGemini CLI
CursorCursor IDE
BlockGoose
JetBrainsJunie
AWSKiro
LettaLetta
Amp Inc. (spun out of Sourcegraph)Amp
SSTOpenCode
All Hands AIOpenHands
SnowflakeCortex Code
DatabricksGenie Code
MistralMistral Vibe
Roo CodeRoo Code
LaravelBoost
ByteDanceTRAE
FactoryFactory

Expanded article table

The pace of adoption surprised even Anthropic. In March 2026 Microsoft published an Azure Skills Plugin built on the standard so that the same package could light up experiences in GitHub Copilot, Copilot CLI, Claude Code, and other supporting tools.[34] Coverage in The New Stack and AI Business compared the open-standard move with Anthropic's success in turning MCP into a de facto standard.[14][17]

The Linux Foundation's Agentic AI Foundation (AAIF) was formed December 9, 2025 with founding contributions from Anthropic, Block, and OpenAI: MCP, Block's goose, and OpenAI's AGENTS.md.[32] Agent Skills initially stayed outside it at agentskills.io. On October 22, 2026, Anthropic donated Agent Skills to the AAIF, one of four new projects announced that day alongside OpenAI's Open Responses, Hugging Face's OpenEnv, and Alibaba's OpenSandbox.[36]

How do Skills compare to Custom GPTs and Gemini Gems?

Every major model provider now ships some form of customization layer. Skills, Custom GPTs, and Google's Gemini Gems all share the goal of packaging reusable AI behavior, but the implementations differ sharply:[12][13]

FeatureClaude SkillsOpenAI Custom GPTsGemini GemsCursor RulesGitHub Copilot Custom Instructions
ProviderAnthropicOpenAIGoogleCursorGitHub
StorageFolder of Markdown plus optional scriptsHosted record on OpenAI serversHosted record on Google servers.cursor/rules/*.mdc files in the repo.github/copilot-instructions.md in the repo
Instruction limitNone (multi-page Markdown)About 8,000 charactersAbout 4,000 charactersNoneNone
Executable codeYes (any language available in the sandbox)Limited to Code Interpreter and Actions (HTTP)Limited to built-in Google toolsNoneNone
External APIsVia the user's environment or scriptsVia Actions (OpenAPI specs)Limited to Google integrationsNoneNone
Discovery modelDescription-based progressive disclosureAll instructions loaded into every conversationAll instructions loaded into every conversationPath-glob matchAlways loaded for repo
SharingGit repo, ZIP, plugin marketplace, public catalogsGPT Store with revenue sharingLink sharing (since September 2025), no public storeRepo onlyRepo only
Cross-vendor portabilityYes, supported by Codex, Cursor, Gemini CLI, GitHub Copilot, and more after the December 2025 open-standard releaseNo (OpenAI-only)NoNo (Cursor-specific)No (GitHub-specific)
ActivationAutomatic, based on description matchManual selection, then auto within the chatManual selection per chatAuto, based on file pathAlways-on for repo
Free tierYes (Skills, including Anthropic's document Skills, are available on the Free claude.ai plan as of November 2026)Yes (since 2024)YesYes (Cursor free tier)Yes (Copilot free tier)

Expanded article table

How do Claude Skills differ from MCP?

Another useful contrast is between Skills and the Model Context Protocol, which is also an Anthropic-led effort. MCP is a protocol for exposing live tools and data sources to a model through a server. Skills are a static format for procedural knowledge. The two are complementary: a Skill might explain how a workflow should be carried out, while an MCP server provides the live data the workflow operates on.[5][14]

AspectClaude SkillsModel Context Protocol
Primary purposeTeach Claude how to perform a taskConnect Claude to external systems
FormatFolder of Markdown and scriptsProtocol specification with stdio or HTTP servers
Token cost when idleRoughly 100 tokens per Skill (description only)Often thousands of tokens for tool listings
Implementation effortWrite a Markdown fileBuild and host a server in a supported language
StateStateless across sessionsCan hold connection state and credentials
Vendor neutralityOpen standard since December 2025; AAIF project since October 2026Vendor-neutral by design (donated to the Linux Foundation's AAIF in December 2025)
DistributionFolder copy or plugin marketplaceServer URL plus auth credentials
DiscoverabilityDescription-based progressive disclosureAll tools loaded at session start

Expanded article table

Anthropic frames the relationship between the various customization features as a stack: a complete Claude agent configuration combines a system prompt, a set of Skills, persistent memory (where supported), and tools (whether built-in, MCP-connected, or invoked through Skill scripts). Skills sit between the always-on system prompt and the per-call tools, providing on-demand domain knowledge.[3][14] Willison put the practical case bluntly: GitHub's official MCP consumes "tens of thousands of tokens," while a Skill that wraps the same workflow as a CLI invocation costs only the description until it triggers.[5]

Skills paired with MCP

By early 2026 the dominant production pattern was to combine Skills with MCP servers rather than choose between them. Anthropic summarizes the split as MCP handling connectivity and Skills handling expertise: with a Skill, Claude can follow your playbook, while MCP tells it where the data lives.[26] One Skill can orchestrate multiple MCP servers. A competitive-analysis Skill might call the Google Drive MCP server for internal notes, a GitHub MCP server for competitor repositories, and a web-search MCP server for market data, then assemble the result using prompts in SKILL.md. Anthropic's December 2025 guidance recommended pairing the two: let MCP handle connectivity, and let Skills handle presentation, sequencing, and workflow logic.[26]

QuestionUse a SkillUse an MCP server
Are you teaching the model how to think about a problem?YesNo
Are you connecting the model to live data or systems?NoYes
Does the workflow change rarely and live in Git?YesEither
Does the resource require credentials, OAuth, or session state?NoYes
Is the cost of always-on tool listings unacceptable?Yes (Skills idle at about 100 tokens)No (full tool list always loaded)
Should the result be portable to Codex, Cursor, and Copilot?Yes (open standard)Yes (open protocol)

Expanded article table

Where can you use Claude Skills?

Skills behave slightly differently depending on the surface they run on. The overall behavior is the same (a folder, progressive disclosure, on-demand load), but each surface has its own constraints:[6][15]

SurfacePre-built SkillsCustom SkillsNetwork accessSharingDistribution
Claude.aiYes (auto, no setup)Yes (Customize > Skills, ZIP upload)Variable per user/admin settingsPrivate by default; Team/Enterprise sharing, org publishing, and owner provisioningManual upload or unified directory
Claude APIYes (skill_id reference)Yes (/v1/skills endpoint)None (sandbox is offline)Workspace-wideAPI endpoint
Claude CodeSelected bundled SkillsYes (filesystem)Same as user's machinePersonal, project, or pluginFilesystem or plugin marketplace
Claude Agent SDKInherited from API or Claude Code sessionInheritedInheritedInheritedWraps Claude Code CLI
Claude CoworkInherited from Claude.ai workspace plus the eleven official Cowork pluginsYes (workspace and admin-provisioned)Same as host machinePlugin- or admin-managedUnified directory or plugin install

Expanded article table

Use in Claude Code

Claude Code discovers Skills in three locations beyond the enterprise scope:[15]

  • Personal Skills. Files placed in ~/.claude/skills/ are available to all of a user's Claude Code projects.
  • Project Skills. Files placed in .claude/skills/ inside a Git repository are available to anyone who clones the project. This makes Skills checkable into version control alongside source code.
  • Plugin-bundled Skills. Skills installed through the Claude Code plugin marketplace appear automatically. A team can publish a private marketplace and install a curated set of Skills across the organization with one command.

In Claude Code, Skills run with the same network and filesystem access as any other program on the user's machine, so they can call external APIs, modify files in the working directory, and execute long-running scripts. This is more permissive than Claude.ai or the API, where Skills run inside a sandboxed virtual machine.[6]

Claude Code merged the older .claude/commands/ directory into the Skills format in January 2026 (version 2.1.3):[38] a file at .claude/commands/deploy.md and a Skill at .claude/skills/deploy/SKILL.md both create the slash command /deploy and behave the same way.[15] Skills add features that plain commands lack: a directory for supporting files, frontmatter to control whether the user or the model invokes the command, and the ability to bundle scripts that Claude can execute. The disable-model-invocation: true flag lets a team build a /deploy Skill that the user must trigger explicitly, so Claude does not decide on its own that the code looks ready to ship.

Several Skill conventions have emerged among Claude Code users. A humanizer Skill for editing AI-generated text, a security-review Skill that audits a pending pull request, and a commit-message Skill that formats Git commit messages are commonly cited examples. The awesome-claude-skills GitHub list curates community Skills focused on developer workflows, and the cross-vendor awesome-agent-skills repository tracks more than a thousand Skills confirmed to work in both Claude Code and the major non-Anthropic CLIs.[9][10][29]

Plugin system relationship

Claude Code introduced a plugin system in October 2025, around the same time as Skills. Plugins are larger packages that can include Skills, slash commands, hooks, MCP server registrations, and configuration. Skills are one of several content types that a plugin can ship.[7][15]

The relationship can be summarized as follows:

ConceptScopeContains
SkillA single capability (one folder with SKILL.md)Instructions, optional scripts, optional reference files
PluginA bundle of related capabilitiesOne or more Skills, slash commands, hooks, settings, MCP server registrations
MarketplaceA registry of pluginsMultiple plugins, each potentially containing many Skills

Expanded article table

A team that publishes a Claude Code marketplace can ship a single plugin called engineering-tooling that bundles a code-review Skill, a release-notes Skill, a slash command for triggering deploys, and an MCP server registration for the team's internal API. Installing the plugin gives every team member the same configuration in one step.[15]

API integration

As of November 2026 the Skills API is generally available and needs no beta header; Skills still run inside the code execution tool:[6][37]

RequirementValue
Beta headerNone (skills-2025-10-02 is still accepted and returns the older beta response shapes)
ToolCode execution tool (code_execution_20250825)
Skill management endpoint/v1/skills
Execution endpointStandard Messages API with container.skills populated
Maximum Skills per request20
Maximum custom Skill upload size30 MB (all files combined, uncompressed)
File outputsReturned with file_id for download via the Files API

Expanded article table

A minimal API request that uses both a built-in Skill and a custom Skill looks like this:

{
  "container": {
    "skills": [
      {"type": "anthropic", "skill_id": "xlsx", "version": "latest"},
      {"type": "custom", "skill_id": "skill_01AbCdEfGhIjKlMnOpQrStUv", "version": "latest"}
    ]
  }
}

Custom Skills uploaded via the API are workspace-wide: every member of the API workspace can use any uploaded Skill. This is different from Claude.ai, where uploads start private to the user and are shared only if the user shares or publishes them.[6][35]

Agent SDK integration

The Claude Agent SDK wraps the Claude Code CLI binary, so any Skills resolvable to the SDK process inherit Claude Code's discovery rules. A developer building an autonomous agent that drafts pull requests, for example, can ship the agent with a .claude/skills/pr-summary directory and trust that the SDK will surface that Skill to Claude on every relevant prompt. This makes Skills the recommended way to package domain knowledge for SDK-based agents, as an alternative to inflating the system prompt or attaching every reference file as context on each turn.

Use in Claude.ai

As of November 2026, Claude.ai exposes Skills under Customize > Skills for users on Free, Pro, Max, Ultra, Team, and Enterprise plans with code execution enabled.[35] Custom Skills are uploaded as ZIP files and stay private to the uploader until shared. Pre-built document Skills (PowerPoint, Excel, Word, PDF) work transparently: if the user asks Claude to make a deck, Claude pulls in the pptx Skill without any explicit setup. Team and Enterprise admins gained centralized organization-wide provisioning on December 18, 2025: from claude.ai/admin-settings/skills they can enable any partner-built Skill for every member of the workspace, and admin-provisioned Skills are enabled by default.[6][23] Later updates let Team and Enterprise users share their own Skills with specific colleagues or groups, or publish them to the organization library, subject to the owner's publishing policy.[35]

Use in Claude Cowork

Claude Cowork, Anthropic's desktop AI agent for macOS and Windows, entered research preview in January 2026 and reached general availability on April 9, 2026.[25] Cowork treats Skills as one of three customization primitives (alongside connectors and plugins) and exposes them through the same unified directory used on Claude.ai. On January 30, 2026, during the research preview, Anthropic open-sourced eleven starter plugins for Cowork, covering areas such as productivity, sales, marketing, finance, enterprise search, legal, customer support, product management, and biology research, plus a meta-plugin for building new plugins; each plugin bundles Skills, connectors, slash commands, and subagents.[33] Users can describe a workflow in a sentence and have Claude assemble a one-shot plugin from existing Skills and connectors. The GA release also added six enterprise features: role-based access controls, group spend limits, usage analytics, OpenTelemetry support, a Zoom MCP connector, and per-tool connector controls.[25]

Enterprise adoption

Anthropic launched Skills with several named enterprise partners and added more in the months after:[1][2]

  • Box integrated Skills so that users can transform stored content into Word, Excel, and PowerPoint documents that follow their organization's brand standards. Yashodha Bhavnani, Head of AI at Box, said: "Skills teaches Claude how to work with Box content. Users can transform stored files into PowerPoint presentations, Excel spreadsheets, and Word documents that follow their organization's standards, saving hours of effort."[1]
  • Notion integrated Skills to make Claude's interactions with Notion more predictable and to reduce prompt engineering for complex tasks. MJ Felix, Product Manager at Notion, said: "With Skills, Claude works seamlessly with Notion, taking users from questions to action faster. Less prompt wrangling on complex tasks, more predictable results."[1]
  • Canva uses Skills to streamline design workflows and maintain brand consistency in generated assets. Anwar Haneef, GM and Head of Ecosystem at Canva, said: "Canva plans to leverage Skills to customize agents and expand what they can do. This unlocks new ways to bring Canva deeper into agentic workflows."[1]
  • Rakuten applied Skills to its finance operations. Yusuke Kaji, General Manager of AI at Rakuten, said: "Skills streamline our management accounting and finance workflows. Claude processes multiple spreadsheets, catches critical anomalies, and generates reports using our procedures. What once took a day, we can now accomplish in an hour."[1]

When the open standard launched in December 2025, partner-built Skills from Atlassian, Cloudflare, Figma, Sentry, Stripe, Vercel, and Zapier joined the existing four, expanding the pool of vetted enterprise integrations.[23] In May 2026, Anthropic launched a vertical-industry push for legal, releasing more than twenty connectors and twelve practice-area plugins (commercial, corporate, employment, privacy, product, regulatory, AI governance, IP, and litigation, among others) packaging Skills for legal workflows.[31]

For Team and Enterprise plans, owners manage Skills from Organization settings > Plugins & skills: they can provision Skills to everyone (enabled by default), scope them to groups through plugins on Enterprise, turn off user-created Skills, and set whether users may publish Skills to the organization.[35]

Are Claude Skills safe to install?

Skills run inside the same code-execution sandbox that Claude uses for any other script. On the Claude API the sandbox has no network access; on Claude.ai the level of network access depends on user and admin settings; in Claude Code, Skills run with the same access as any other process on the user's machine.[6]

Anthropic's official guidance on security includes:[6]

  • Use only Skills you wrote yourself or obtained from a trusted source.
  • Audit all bundled files (SKILL.md, scripts, images, references) before installing a third-party Skill.
  • Be especially careful with Skills that fetch data from external URLs, since fetched content can contain prompt-injection payloads.
  • Treat Skill installation like installing any other piece of software on a production system.
  • Consider least-privilege configurations: a Skill that does not need network access should not be granted it.

Because a malicious Skill can in principle direct Claude to invoke any tool the agent has access to, an audit step before installing community Skills is non-optional. The official documentation puts it plainly: "a malicious Skill can direct Claude to invoke tools or execute code in ways that don't match the Skill's stated purpose."[6] In Claude Code, the project trust dialog protects against this for repository-checked Skills: the allowed-tools field only takes effect after the user accepts the workspace trust prompt for that folder.

Claude Skills data is not eligible for Zero Data Retention: Skill definitions and execution data are retained per Anthropic's standard retention policy.[6]

How do you write a good Skill?

Anthropic publishes a best-practices document for Skill authors. The most repeated points are:[3][6][15]

  • Write a precise description that says both what the Skill does and when Claude should use it. Vague descriptions cause non-activation.
  • Keep one Skill focused on one capability. A Skill that tries to do everything will be triggered for the wrong tasks.
  • Put long reference material into separate files (FORMS.md, REFERENCE.md) and link to them from SKILL.md. This keeps level-2 loading cheap.
  • Use scripts for any operation that is deterministic. Letting the model regenerate code on every call is slower and less reliable than running a checked-in script. Anthropic's engineering blog puts the case in concrete terms: "sorting a list via token generation is far more expensive than simply running a sorting algorithm."[3]
  • Test the Skill with a variety of phrasings to make sure activation is reliable.
  • Version the Skill explicitly when behavior changes, so callers can pin to a known-good version.
  • Keep SKILL.md itself under 500 lines. Move detailed reference material to separate files.[15]
  • Treat the body as standing instructions, not one-time commands, since the rendered content stays in context across turns once activated.

The skill-creator Skill, which ships in the Anthropic skills repository, will walk a user through these steps interactively: it asks about the workflow, generates the folder structure, fills in the YAML frontmatter, and bundles requested resources without manual file editing.[7]

What are the limitations of Claude Skills?

The published documentation lists several limitations and constraints:[6]

  • Limited cross-surface sync for custom Skills. A Skill uploaded on Claude.ai must be uploaded separately to the API, and API Skills do not appear on Claude.ai.[6] Claude Code is the partial exception: Skills enabled on a user's claude.ai account, including organization-provisioned ones, also load in Claude Code and Cowork sessions signed in with that account.[15][35]
  • Sharing on Claude.ai depends on plan. Custom Skills start private to the uploader. On Team and Enterprise plans users can share them with colleagues or groups or publish them to the organization, subject to the owner's policy; individual plans have no organization-wide sharing.[35]
  • API sandbox is offline. Skills running through the Claude API have no network access and cannot install new packages at runtime. Only pre-installed packages in the code-execution container are available.
  • Twenty Skills per request on the API. A single Messages API request can include at most twenty Skills.[37]
  • 30 MB upload limit. A custom Skill uploaded via the API must total 30 MB or less, all files combined and uncompressed.[37]
  • Not eligible for Zero Data Retention. Skill definitions and execution traces are kept under Anthropic's standard retention policy.
  • Activation depends on the description. Skills with weak descriptions may not trigger reliably; this requires authoring care.
  • Trust required for third-party Skills. Because Skills can run scripts, installing one from an untrusted source is equivalent to running an untrusted program.
  • Description budget. Claude Code truncates the combined description and when_to_use text at 1,536 characters in the listing, and the overall character budget for the Skill listing scales at 1% of the model's context window, adjustable with the skillListingBudgetFraction setting.[15]

Reception

Reception in the developer community was strongly positive. Simon Willison's October 16, 2025 post argued that Skills "might be a bigger deal than MCP" because the format is dramatically simpler than the Model Context Protocol and because progressive disclosure solves the context-window cost problem that has dogged MCP servers in production. He predicted what he called a "Cambrian explosion" of community Skills.[5]

Within weeks, third-party marketplaces and curated lists appeared. The awesome-claude-skills GitHub list, the claude-skills-marketplace repository, the tonsofskills.com directory, and the SkillsMP catalog were all live by late 2025, collectively listing thousands of community Skills aimed at engineering, marketing, product, compliance, and executive workflows.[9][10][11] Coverage in InfoQ, Open Data Science, The New Stack, AI Business, and Neowin emphasized the open-standard nature of the format and its potential to become a cross-vendor configuration layer for AI agents.[8][14][16][17][18]

Simon Willison's December 19, 2025 follow-up on the open-standard release pointed out the deliberate minimalism of the specification: brief enough to read in one sitting, lightly defined enough to leave room for vendor-specific extensions like Claude Code's disable-model-invocation field.[21] By mid-2026, more than thirty independent agent products read the same SKILL.md format from the same directory structure.[22]

The scale caught most observers off guard. By early 2026 Vercel's skills.sh indexed tens of thousands of community uploads. Some marketplaces, notably Agensi, sell paid Skills and pay authors a share of each sale (Agensi advertises a 70 percent creator split), bringing storefront economics like the GPT Store to the Skills ecosystem.[28]

Skills succeeded where prior customization attempts (Custom GPTs, Gemini Gems) faltered because Skills are files. There is no upload, no hosting layer, no platform tax, no review process. A Skill is a directory in a Git repository. That choice is what made cross-vendor adoption possible: every agentic tool already knew how to read a folder.

When were Claude Skills released? (Timeline)

DateEventSource
October 16, 2025Public announcement of Agent Skills; available on Claude.ai (Pro, Max, Team, Enterprise), Claude Code, the Claude API, and the Claude Agent SDK[1][2]
October 2025Public release of the anthropics/skills GitHub repository with example Skills and document Skills[7]
October 2025Launch partners Box, Notion, Canva, and Rakuten announce production integrations[1][2]
October 2025Claude Code plugin system enters public beta, with Skills as one supported content type[15]
October 9, 2025Jesse Vincent (obra) publishes the Superpowers Skills framework, which becomes the most-starred community Skill project[30]
January 7, 2026 (Claude Code 2.1.0)Hot reload for Skills, session teleportation, language-specific output[38]
December 18, 2025Anthropic releases the Agent Skills format as an open standard at agentskills.io; the specification is published as the independent agentskills/agentskills GitHub repository; partner directory and org-wide admin provisioning ship for Team and Enterprise[20][21][23]
December 2025Microsoft supports Skills in VS Code and GitHub at the open-standard launch; OpenAI, which had quietly added the format to ChatGPT and Codex CLI earlier in the month, documents Skills for Codex on December 20[20][21]
January 2026Superpowers added to Anthropic's official Claude Code plugin marketplace[30]
January 2026Claude Cowork enters research preview with Skills as a first-class building block[25]
January 30, 2026Anthropic open-sources eleven starter plugins built on Skills for Claude Cowork[33]
Early 2026Vercel's skills.sh directory and other third-party marketplaces collectively list tens of thousands of community Skills[9][10][11]
March 9, 2026Microsoft publishes the Azure Skills Plugin built on the open standard[34]
March 31, 2026Anthropic launches the unified directory at claude.ai/customize for skills, connectors, and plugins[24]
April 9, 2026Claude Cowork reaches general availability on macOS and Windows with enterprise features including role-based access controls, OpenTelemetry support, and a Zoom MCP connector[25]
May 2026Anthropic ships more than twenty connectors and twelve practice-area plugins for legal workflows[31]
Mid-2026More than thirty independent agent products read the same SKILL.md format[22]
October 22, 2026Anthropic donates Agent Skills to the Linux Foundation's Agentic AI Foundation[36]

Expanded article table

Frequently asked questions

What are Claude Skills in one sentence? Claude Skills are folders containing a SKILL.md instruction file plus optional scripts and reference files that Claude loads on demand to specialize at a repeatable task, using progressive disclosure to keep the context window small.[1][3]

When did Anthropic launch Skills? Anthropic announced Agent Skills on October 16, 2025 and made them available the same day across Claude.ai, Claude Code, the Claude API, and the Claude Agent SDK.[1][2]

Are Claude Skills free? Skills carry no extra charge. As of November 2026 they are available on every Claude.ai plan, including Free, with code execution enabled; Anthropic's document Skills (PowerPoint, Excel, Word, PDF) work automatically for all users, and API use follows standard API pricing with the code execution tool.[35][37]

Is the Skills format open source? The Agent Skills format itself is an open standard published at agentskills.io on December 18, 2025, and most Skills in the anthropics/skills repository are licensed under Apache 2.0. The exception is the four document Skills, which are source-available rather than open source.[7][20]

Do Skills work outside Claude? Yes. After the December 2025 open-standard release, more than thirty agent products read the same SKILL.md files, including GitHub Copilot, VS Code, Cursor, Google's Gemini CLI, and OpenAI's Codex CLI.[20][22]

How are Skills different from MCP? Skills teach Claude how to perform a task as static Markdown; the Model Context Protocol connects Claude to live external systems through a server. They are complementary, and the dominant 2026 production pattern pairs a Skill (the playbook) with one or more MCP servers (the data sources).[5][26]

Continue reading

More from See also

Topics:AI Tools & ProductsAnthropic

References

  1. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12"Introducing Agent Skills." Anthropic. October 16, 2025. anthropic.com/...skills (redirects to claude.com/...skills)
  2. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9"Introducing Agent Skills." Claude (Anthropic). October 16, 2025. claude.com/...skills
  3. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12"Equipping agents for the real world with Agent Skills." Anthropic Engineering. October 2025. anthropic.com/...-the-real-world-with-agent-skills
  4. ^1 ^2 ^3Lee, Hanchung. "Claude Agent Skills: A First Principles Deep Dive." October 26, 2025. leehanchung.github.io/...claude-skills-deep-dive
  5. ^1 ^2 ^3 ^4 ^5Willison, Simon. "Claude Skills are awesome, maybe a bigger deal than MCP." simonwillison.net. October 16, 2025. simonwillison.net/...claude-skills
  6. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17 ^18 ^19 ^20 ^21"Agent Skills." Claude API Documentation. platform.claude.com/...overview (accessed May 2026).
  7. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9Anthropic. "anthropics/skills." GitHub repository. github.com/...skills (accessed May 2026).
  8. ^1 ^2"Anthropic Introduces Skills for Custom Claude Tasks." InfoQ. October 2025. infoq.com/...anthropic-claude-skills
  9. ^1 ^2 ^3 ^4Travisvn. "awesome-claude-skills." GitHub. github.com/...awesome-claude-skills (accessed May 2026).
  10. ^1 ^2 ^3 ^4Rezvani, Alireza. "claude-skills." GitHub. github.com/...claude-skills (accessed May 2026).
  11. ^1 ^2 ^3Longshore, Jeremy. "claude-code-plugins-plus-skills." GitHub. github.com/...claude-code-plugins-plus-skills (accessed May 2026).
  12. ^"Claude Skills vs Gemini Gems vs ChatGPT GPTs: Which One Wins?" FindSkill.ai. findskill.ai/...lls-vs-gemini-gems-vs-chatgpt-gpts (accessed May 2026).
  13. ^"Detailed Comparison: Custom GPTs in ChatGPT vs Claude, Gemini." AI Fire. aifire.co/...stom-gpts-in-chatgpt-vs-claude-gemini (accessed May 2026).
  14. ^1 ^2 ^3 ^4"Anthropic Launches Skills Open Standard for Claude." AI Business. December 18, 2025. aibusiness.com/...ches-skills-open-standard-claude
  15. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17 ^18"Extend Claude with skills." Claude Code Documentation. code.claude.com/...skills (accessed May 2026).
  16. ^"Anthropic Introduces Agent Skills: Making Claude Smarter, Faster, and More Specialized." Open Data Science. October 2025. opendatascience.com/...faster-and-more-specialized
  17. ^1 ^2"Agent Skills: Anthropic's Next Bid to Define AI Standards." The New Stack. December 18, 2025. thenewstack.io/...-next-bid-to-define-ai-standards
  18. ^"Anthropic adds composable 'Skills' to Claude for specialized workflow support." Neowin. October 2025. neowin.net/...ude-for-specialized-workflow-support
  19. "Claude Skills Marketplace: The Essential Plugin Hub for Developers." BrightCoding. April 26, 2026. blog.brightcoding.dev/...plugin-hub-for-developers
  20. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8"Anthropic launches enterprise Agent Skills and opens the standard, challenging OpenAI in workplace AI." VentureBeat. December 18, 2025. venturebeat.com/...t-skills-and-opens-the-standard
  21. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8Willison, Simon. "Agent Skills." simonwillison.net. December 19, 2025. simonwillison.net/...agent-skills
  22. ^1 ^2 ^3 ^4 ^5"Agent Skills: Overview." agentskills.io. agentskills.io (accessed May 2026).
  23. ^1 ^2 ^3 ^4 ^5 ^6"Skills for organizations, partners, and the ecosystem." Claude (Anthropic). December 18, 2025. claude.com/...organization-skills-and-directory
  24. ^1 ^2 ^3 ^4"Browse skills, connectors, and plugins in one directory." Claude Help Center. March 31, 2026. support.claude.com/...and-plugins-in-one-directory
  25. ^1 ^2 ^3 ^4"Anthropic Announces Claude Cowork." InfoQ. January 2026. infoq.com/...claude-cowork; "Anthropic takes Claude Cowork out of preview and straight into the enterprise." The New Stack. April 9, 2026. thenewstack.io/...and-straight-into-the-enterprise
  26. ^1 ^2 ^3"Extending Claude's capabilities with skills and MCP servers." Claude (Anthropic). December 19, 2025. claude.com/...capabilities-with-skills-mcp-servers
  27. ^1 ^2 ^3 ^4 ^5 ^6 ^7"Specification." Agent Skills. agentskills.io/specification (accessed May 2026).
  28. ^"How AI Agent Skill Marketplaces Work: Developer's Guide (2026)." Agensi. agensi.io/...how-ai-agent-skill-marketplaces-work (accessed May 2026).
  29. ^1 ^2VoltAgent. "awesome-agent-skills." GitHub. github.com/...awesome-agent-skills (accessed May 2026).
  30. ^1 ^2 ^3 ^4Vincent, Jesse (obra). "superpowers." GitHub. github.com/...superpowers (accessed May 2026).
  31. ^1 ^2"Anthropic Goes All-In on Legal, Releasing More Than 20 Connectors and 12 Practice-Area Plugins for Claude." LawSites. May 2026. lawnext.com/...12-practice-area-plugins-for-claude
  32. ^"Linux Foundation Announces the Formation of the Agentic AI Foundation (AAIF), Anchored by New Project Contributions Including Model Context Protocol (MCP), goose and AGENTS.md." Linux Foundation press release. December 9, 2025. linuxfoundation.org/...f-the-agentic-ai-foundation
  33. ^1 ^2Lardinois, Frederic. "Anthropic brings plugins to Cowork." The New Stack. January 30, 2026. thenewstack.io/anthropic-brings-plugins-to-cowork
  34. ^1 ^2Harris, Chris. "Announcing the Azure Skills Plugin." Microsoft All things Azure blog. March 9, 2026. devblogs.microsoft.com/...-the-azure-skills-plugin
  35. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9"Use skills in Claude" and "Provision and manage skills for your organization." Claude Help Center. support.claude.com/...12512180-use-skills-in-claude ; support.claude.com/...skills-for-your-organization (accessed November 2026).
  36. ^1 ^2 ^3"Agentic AI Foundation Announces Four New Projects: Agent Skills, Open Responses, OpenEnv, and OpenSandbox." Linux Foundation press release. October 22, 2026. linuxfoundation.org/...ses-openenv-and-opensandbox
  37. ^1 ^2 ^3 ^4 ^5 ^6"Using Agent Skills with the API." Claude API Documentation. platform.claude.com/...skills-guide (accessed November 2026).
  38. ^1 ^2 ^3Anthropic. "Claude Code changelog" (versions 2.1.0, January 7, 2026, and 2.1.3, January 9, 2026). GitHub. github.com/...CHANGELOG.md

Improve this article

Add missing citations, update stale details, or write a clearer explanation. Anyone can edit: changes go live when you publish, and every edit is checked by an editor.

10 revisions · v10 · 9,290 words · full history

Fact-checks are independent of edits: a reviewer re-verifies the article against its sources and stamps the date. How we verify

Research and drafting on this wiki are AI-assisted, under named human editorial standards. How AI is used here

Reviewer note: Independent fact-check of the full article against its cited sources (210 claims, 44 sources fetched); 31 material + 38 minor fixed

Cite this page: AI Wiki. "Claude Skills." aiwiki.ai, updated 22 Jun 2026, fact-checked 4 Nov 2026. CC BY 4.0. https://aiwiki.ai/wiki/claude_skills

Edit article