Claude Cowork Skills and Plugins: What Each Component Does

Claude Cowork plugins package reusable instructions and integrations for a role or workflow. A plugin can include skills, connectors, slash commands, and subagents. You install and share the plugin as a unit; Claude uses the components inside it to carry out specific work.

The terms are easy to confuse because they overlap in the interface. The practical distinction is scope: a skill teaches Claude how to perform one kind of task, while a plugin assembles several capabilities into a reusable toolkit.

Skills, plugins, commands, connectors, and subagents

Component	Purpose	Use it when
Skill	Reusable instructions, references, and optional resources for a focused task	Claude should follow a repeatable method, style, checklist, or output format
Plugin	A distributable package containing one or more Cowork components	A team or job function needs a coordinated toolkit
Slash command	An explicit action shown in the `/` menu	You want to start a known workflow on demand
Connector	Provides controlled access to an external service or data source, commonly through MCP	The workflow must read or act on information outside Claude
Subagent	A specialized agent that handles a bounded part of a larger task	Complex work benefits from separate roles, instructions, tools, or context

What a skill does

A skill captures a procedure that would otherwise need to be explained repeatedly. For example, a contract-review skill might define which clauses to inspect, how to report risks, which issues require a lawyer, and the expected output table. A writing skill might contain a company's voice rules and editing checklist.

The central file is commonly named SKILL.md. Its description helps Claude decide when the skill applies, while its body contains the workflow. A well-designed skill is narrow enough to trigger predictably and complete enough that another user can understand its assumptions.

A skill can be selected from the / menu after it has been installed through a plugin. Depending on the product surface and the skill's description, Claude may also use relevant instructions as part of a broader task. Do not assume automatic selection will always be perfect; name the skill explicitly when consistency matters.

What a plugin adds

A plugin is the container you install, customize, or distribute. A marketing plugin, for example, could combine campaign-planning and content-brief skills, a command for creating a weekly report, CRM and analytics connectors, and a subagent that researches competitors.

A minimal plugin may contain only one skill. More advanced plugins can add manifests, connector configuration, multiple agents, and supporting files. Simple instruction-only packages are largely text-based, but it is misleading to say that every plugin takes only minutes or requires no technical work. Secure connectors, error handling, testing, and deployment governance can require substantial effort.

How the components work together

The user installs a plugin for a role or workflow.
The user starts a task or chooses an installed skill from the / menu.
The skill supplies the procedure and expected result.
A connector retrieves or updates approved external data when the task requires it.
A subagent can take responsibility for a bounded research, analysis, or drafting step.
Cowork returns the combined result for review.

Not every workflow needs all six steps. If a task only requires consistent formatting, use a skill. Add a connector only when outside data is necessary, and add subagents only when separation genuinely improves the work.

Install a plugin in Claude Cowork

Open the Cowork tab in Claude.
Open Customize from the left sidebar.
Select the Plugins tab, then choose Browse plugins.
Review the plugin's source, skills, connectors, and permissions.
Select Install.
In a Cowork task, type / or use the plus button to see skills from installed plugins.

Anthropic also supports adding approved marketplaces and uploading a custom plugin file. Team or Enterprise owners can distribute organization-managed plugins; required plugins may not be removable by individual users. Availability and controls can vary by plan and administrator settings, so consult Anthropic's current plugin instructions.

Customize an installed plugin

In Cowork, open the installed plugin and select Customize. Claude starts a task that helps adapt the package. Useful changes include:

replacing generic terminology with your team's vocabulary;
adding approved templates, review gates, and escalation rules;
removing components your team does not use;
connecting only the required systems and data scopes; and
adding examples that demonstrate acceptable and unacceptable output.

Keep organization-specific values separate from the reusable core where possible. Version the plugin, document its owner, and test changes before distributing them. Organization-managed plugins cannot be edited by ordinary users, which helps keep shared workflows consistent.

Anthropic's plugin marketplaces

Anthropic provides built-in marketplaces, including a Knowledge Work marketplace, plus additional Anthropic-curated sources for areas such as legal, financial services, and life sciences. The open-source Knowledge Work Plugins repository includes role-oriented examples for Cowork and Claude Code.

The catalog and number of components change over time. A fixed claim such as �15 plugins and more than 85 skills� can quickly become wrong; browse the live marketplace or repository instead of relying on an old count.

When to use each component

Need	Best starting point
Apply the same writing or review method repeatedly	Skill
Launch a repeatable workflow from the menu	Slash command or invocable skill
Bundle a complete toolkit for a department	Plugin
Read data from a CRM, drive, database, or API	Connector with the minimum required permissions
Delegate a bounded specialist task	Subagent
Call Claude directly from your own application	Claude API, not a Cowork plugin alone

For a related developer-oriented comparison, see TipsMake's guide to Claude Code skills and plugins. If you need a backend integration rather than a Cowork customization, start with the Claude API guide.

Security checks before installation

A plugin can influence Claude's behavior and may gain access to connected services, so treat third-party packages like software�not like harmless prompt templates.

Confirm the publisher and source repository.
Read the skills, agent instructions, manifest, and connector definitions.
Grant the smallest practical data and action scope.
Keep credentials in the supported secret store; never paste them into a skill file.
Test in a non-production workspace with sample data.
Check whether the workflow can send data to another service or execute consequential actions.
Review updates before accepting material permission or behavior changes.

Anthropic says Enterprise organizations can enable skill scanning, which blocks detected malicious content and warns about potential risk. Scanning is an additional control, not a substitute for source review, access limits, and human approval of high-impact actions.

A sensible first project

Start with one low-risk skill, such as turning meeting notes into a fixed internal template. Define the trigger, required input, output structure, and review checklist. Test it on varied examples, then package it as a plugin only when multiple people need to install or customize it. Add connectors and subagents later, once the basic workflow is reliable.

This staged approach makes errors easier to diagnose and keeps permissions proportional to the actual task. It also produces a clearer, more maintainable plugin than trying to automate an entire department in the first version.