Claude Skills, Cowork, and Plugins: The Complete Guide for Marketing Teams

Claude released Cowork and plugins in January 2026. Most marketers still treat Claude like a chatbot. The teams pulling ahead use Claude as an autonomous operator that executes complete workflows while they focus on strategy.

This guide breaks down the entire Claude product ecosystem: Claude.ai, Claude Desktop, Claude Code, Skills, MCP, Cowork, and plugins. You will learn what each component does, how they connect, and how to build marketing workflows that run without your constant input.

TL;DR

Claude offers three interfaces: Claude.ai (browser), Claude Desktop (native app with local file access), and Claude Code (terminal for developers). Skills are pre-written instructions that teach Claude repeatable workflows. MCP connects Claude to external tools like HubSpot and Slack. Cowork brings Claude Code’s autonomous execution to non-developers. Plugins bundle skills, commands, connectors, and sub-agents into role-specific packages that turn Claude into a marketing specialist.

Key Takeaways

• Claude.ai handles conversational tasks while Cowork executes multi-step workflows autonomously
• Skills are markdown files with instructions that Claude loads automatically when relevant
• MCP provides tool connections; skills provide the methodology for using those tools
• Cowork plugins bundle four components: skills, commands, connectors, and sub-agents
• Sub-agents enable parallel task execution with isolated context windows
• Marketing teams save 10-15 hours per week by replacing manual prompting with plugin-based workflows
• Plugins are file-based (markdown and JSON) with no code required to customize

What Are the Three Claude Products?

Claude operates through three distinct interfaces. Each serves different use cases and user types.

Claude.ai (Web and Mobile)

Claude.ai is the browser-based chat interface at claude.ai plus iOS and Android apps. Users type messages, upload files, and receive responses in a conversational format.

Core features include Artifacts (a dedicated panel for code, documents, and React components), Projects (workspaces with 200K context windows), web search, and 75+ managed MCP connectors. Skills upload via Settings > Capabilities > Skills.

Claude.ai works best for content drafting, research, brainstorming, and document analysis. The conversational format limits autonomous execution but excels at iterative collaboration.

Claude Desktop Application

Claude Desktop is a native Mac and Windows application providing the same features as Claude.ai plus local MCP server support. The key difference: Claude Desktop accesses files on your machine without uploading them to the cloud.

Configuration lives at ~/Library/Application Support/Claude/claude_desktop_config.json on Mac. Local MCP servers enable database connections, script execution, and file system operations that cloud-based Claude.ai cannot perform.

Claude Desktop works best for processing sensitive data, large file operations, and workflows requiring local tool access. It includes the Cowork tab for autonomous task execution.

Claude Code (Terminal Tool)

Claude Code is a command-line tool for agentic coding and automation. Run claude from any project directory. Claude Code reads your codebase, executes shell commands, edits files, runs tests, and commits changes.

Claude Code operates in a working directory and automatically reads CLAUDE.md files for project context. It supports sub-agents, hooks, and full shell access. Pricing ties to Claude Pro/Max subscriptions or API token usage.

Claude Code works best for developers building automation, technical marketing operations, and multi-step workflows requiring code execution. Non-developers should use Cowork instead.

Action item: Download Claude Desktop from claude.ai/download. Open Settings > Capabilities > Skills to see where you will upload custom skills later.

What Is the Difference Between Skills, MCP, and Plugins?

These three terms cause the most confusion. Here is the clear distinction.

Skills: Pre-Written Instructions

A skill is a folder containing instructions that teach Claude repeatable workflows. The folder must include one required file: SKILL.md. No code, no infrastructure, no build steps.

Without a skill, you explain your brand voice, email format, and campaign structure every conversation. With a skill, Claude reads your instructions automatically and follows your process.

Skills work everywhere: Claude.ai, Claude Desktop, and Claude Code. Upload skills via Settings > Capabilities > Skills (zipped folder) or place them in ~/.claude/skills/ for Claude Code.

MCP: Tool Connections

MCP (Model Context Protocol) connects Claude to external systems. Think of MCP as plumbing that gives Claude access to Salesforce, Notion, Google Analytics, and other platforms.

Claude.ai includes 75+ managed MCP connectors you enable with one click. Claude Desktop and Claude Code support local MCP servers you configure yourself for custom integrations.

MCP defines what Claude accesses. Skills define how Claude uses those tools.

Plugins: Bundled Packages

Plugins bundle four components into a single package: skills (domain expertise), commands (slash triggers), connectors (MCP integrations), and sub-agents (parallel workers). Install one plugin and Claude becomes a specialist for a specific role.

Plugins launched January 30, 2026 for Cowork. They work in Claude Desktop’s Cowork tab and are compatible with Claude Code. Anthropic open-sourced 11 plugins covering sales, marketing, finance, legal, and other functions.

Action item: Enable one MCP connector in Claude.ai. Go to Settings > Connectors and connect Slack or your project management tool. Test by asking Claude to “check my recent Slack messages.”

How Do Skills Work Technically?

Skills follow a specific structure that enables progressive disclosure. Claude loads only what it needs, when it needs it.

Skill Folder Structure

Every skill folder contains these components:

• SKILL.md (required): Instructions in Markdown with YAML frontmatter
• scripts/ (optional): Executable code (Python, Bash)
• references/ (optional): Documentation loaded as needed
• assets/ (optional): Templates, fonts, icons used in output

The folder name uses kebab-case only: campaign-brief-generator works, Campaign Brief Generator breaks.

SKILL.md File Format

The SKILL.md file has two parts: YAML frontmatter and markdown body.

The frontmatter requires two fields: name (kebab-case identifier) and description (what it does and when to use it). Claude reads the description to decide whether to load the skill.

---
name: email-sequence-writer
description: Creates email nurture sequences. Use when user mentions email sequence, drip campaign, or nurture emails.
---

The markdown body contains your full instructions: step-by-step workflows, quality standards, examples, and output formats. Claude loads this content only when it determines the skill applies to the current task.

Progressive Disclosure System

Skills use three levels of loading to minimize token usage:

Level 1 (YAML frontmatter): Always loaded in Claude’s system prompt. Contains name and description only. This tells Claude when to activate the skill.

Level 2 (SKILL.md body): Loaded when Claude determines the skill applies. Contains full instructions, examples, and quality standards.

Level 3 (Linked files): References in references/ and scripts/ loaded only when Claude navigates to them during execution.

This system keeps Claude’s context clean. A marketing team with 20 skills installed does not load all 20 into every conversation.

Action item: Create your first skill folder. Make a folder called brand-voice-guide/ containing one file: SKILL.md. Add frontmatter with name and description, then paste your brand voice guidelines in the body.

What Does a Marketing Skill Look Like?

Here is a complete skill example for competitive analysis. Copy this structure for your own workflows.

Complete Skill Example

Create a folder called competitive-analyst/ with this SKILL.md file:

---
name: competitive-analyst
description: Analyzes competitor positioning, messaging, and campaigns. Use when user asks about competitors, competitive analysis, or market positioning.
license: MIT
metadata:
  author: yfxmarketer
  version: 1.0.0
---

The markdown body follows the frontmatter:

# Competitive Analysis Workflow

## When to Use

Activate this skill when the user mentions competitor names, asks about market positioning, requests competitive intelligence, or needs campaign differentiation.

## Process

1. Identify the competitor or competitors to analyze
2. Research their current messaging across website, ads, and social
3. Document their positioning statement and value propositions
4. Identify 3 key differentiators versus our offering
5. Note gaps or weaknesses in their approach
6. Recommend positioning angles for our campaigns

## Output Format

Deliver analysis as a structured brief with these sections:

- Competitor Overview (company, market position, target audience)
- Messaging Analysis (headlines, value props, proof points)
- Campaign Examples (recent ads, landing pages, email samples)
- Differentiation Opportunities (gaps we exploit)
- Recommended Angles (3 positioning options with rationale)

## Quality Standards

- Include specific examples with URLs when available
- Quote actual competitor copy, do not paraphrase
- Provide actionable recommendations, not observations
- Keep total brief under 1,500 words

This skill triggers whenever you mention competitors. Claude follows the process, uses the output format, and meets quality standards without you specifying them each time.

Action item: Adapt this template for one of your workflows. Replace the process steps with your actual methodology. Upload the zipped folder to Settings > Capabilities > Skills.

What Is Claude Cowork?

Cowork launched January 12, 2026 as Claude Code for non-developers. It brings autonomous execution to knowledge workers through the Claude Desktop app.

How Cowork Differs from Chat

Chat operates on a simple loop: you provide input, Claude generates output, you evaluate and iterate. Each exchange is discrete and requires your attention.

Cowork operates differently. You describe an outcome, approve Claude’s plan, and step away. Claude executes multi-step tasks autonomously, coordinates parallel workstreams, and delivers finished work. You return to completed deliverables, not conversations.

The interface lives in Claude Desktop as a tab alongside Chat. Click Cowork, grant folder access, describe your task, approve the plan, and let Claude work.

What Cowork Executes

Cowork handles any task achievable through file operations and tool access:

• File organization: Sort and rename hundreds of downloads by type and date
• Data extraction: Create expense spreadsheets from receipt screenshots
• Document creation: Draft reports from scattered notes across multiple files
• Research synthesis: Read and summarize findings from large document sets
• Batch processing: Apply consistent transformations across many files

Claude Code users discovered they could do all this in the terminal. Cowork removes the terminal requirement while keeping the autonomous execution model.

Technical Architecture

Cowork runs on the same Claude Agent SDK that powers Claude Code. Tasks execute in a virtual machine environment on your computer. Claude reads designated folders, coordinates sub-agents, and writes outputs directly to your file system.

The Claude Desktop app must remain open during task execution. Internet connection stays active throughout. Cowork respects your network egress permissions set in Claude Desktop settings.

Action item: Open Claude Desktop and click the Cowork tab. Grant access to your Downloads folder. Ask Claude to “organize files by type and create subfolders.” Watch Claude make a plan and execute it.

How Do Cowork Plugins Enable Orchestration?

Plugins turn Claude into a role-specific specialist by bundling four components that work together.

The Four Plugin Components

Skills define methodology. They encode domain expertise, best practices, and step-by-step workflows. Claude loads skills automatically when relevant. A sales skill teaches Claude your qualification criteria. A marketing skill teaches your campaign planning process.

Commands trigger actions. Slash commands like /marketing:campaign-brief invoke specific workflows. One command triggers an entire multi-step process. Users invoke one command; Claude executes complete workflows.

Connectors provide tool access. The .mcp.json file configures which external systems Claude accesses. A marketing plugin connects to HubSpot, Amplitude, Notion, and Slack. Claude pulls and pushes data across all connected systems.

Sub-agents execute parallel tasks. Specialized Claude instances with isolated context windows handle discrete work. The main Claude (orchestrator) spawns sub-agents, waits for results, and compiles outputs.

Plugin File Structure

Every plugin follows this structure:

marketing/
├── .claude-plugin/plugin.json   # Plugin manifest
├── .mcp.json                    # Tool connections
├── commands/                    # Slash commands
│   ├── campaign-brief.md
│   └── content-calendar.md
└── skills/                      # Domain knowledge
    ├── brand-voice.md
    └── campaign-planning.md

All files are markdown and JSON. No code compilation, no infrastructure deployment, no build steps. Edit the files directly to customize behavior.

Orchestration Flow Example

When you type /marketing:campaign-brief "Q2 Product Launch":

1. Command triggers the campaign-brief workflow
2. Workflow activates the campaign-planning skill
3. Claude spawns research sub-agent to analyze competitors (runs in parallel)
4. Claude spawns content sub-agent to draft messaging (runs in parallel)
5. Claude calls Amplitude MCP to pull past campaign performance
6. Claude calls HubSpot MCP to retrieve audience segment data
7. Orchestrator waits for sub-agents to complete
8. Orchestrator compiles all inputs using skill template
9. Output delivered as formatted campaign brief document

Total time: 2-5 minutes. Manual equivalent: 2-4 hours.

Action item: Install the marketing plugin in Cowork. Go to Plugins, find “marketing,” and click Install. Test with /marketing:campaign-brief for a recent project.

How Do Sub-Agents Enable Parallel Execution?

Sub-agents are the core orchestration mechanism. They solve the context pollution problem that degrades Claude’s performance on complex tasks.

The Context Problem

Long conversations fill Claude’s context window with irrelevant information. By message 50, earlier context matters less than recent exchanges. Quality degrades as conversations grow.

Sub-agents solve this by isolating work. Each sub-agent starts with fresh context, executes its specific task, returns results, and terminates. The orchestrator receives clean outputs without inheriting conversation history.

Built-In Sub-Agent Types

Cowork and Claude Code include three built-in sub-agents:

Explore uses Haiku (fast, cheap) with read-only tools. It scans files, searches code, and maps directory structures. Purpose: rapid information gathering without modification risk.

Plan uses Sonnet with read-only exploration tools. It researches before creating plans. Purpose: gather context without spawning infinite agent chains.

General-purpose uses Sonnet with all tools. It handles complex multi-step tasks requiring both exploration and action. Purpose: work that needs reasoning, modification, and verification.

Custom Sub-Agents in Plugins

Plugins define custom sub-agents as markdown files with YAML frontmatter:

---
name: lead-qualifier
description: Qualifies leads based on ICP criteria. Use when new leads enter pipeline.
tools: Read, Grep, Bash
model: haiku
---

You are a lead qualification specialist.

When invoked:
1. Pull lead data from CRM
2. Check against ICP criteria in /docs/icp.md
3. Score lead on 1-10 scale
4. Flag missing information
5. Recommend next action

Return structured qualification report.

The orchestrator spawns this sub-agent when it encounters lead qualification tasks. The sub-agent runs with restricted tools (Read, Grep, Bash only), uses a cheaper model (Haiku), and returns results to the orchestrator.

Parallel Execution Pattern

Complex tasks spawn multiple sub-agents simultaneously:

User: "Prepare Q2 campaign launch package"

Orchestrator receives task
    │
    ├─► Research Sub-Agent (Haiku, read-only)
    │   • Competitor analysis
    │   • Market trends
    │   • Returns: research brief
    │
    ├─► Content Sub-Agent (Sonnet, write access)
    │   • Email sequences
    │   • Ad copy variants
    │   • Returns: draft assets
    │
    └─► Analytics Sub-Agent (Haiku, read-only)
        • Past performance data
        • Audience insights
        • Returns: metrics report

Orchestrator waits for all three
Orchestrator compiles into campaign package
User receives finished deliverable

Three sub-agents work simultaneously. The orchestrator adds no overhead while waiting. Total time equals the slowest sub-agent plus compilation.

Action item: In Cowork, ask Claude to “research three competitors and draft positioning for each.” Watch Claude spawn parallel sub-agents in the progress indicator.

Which Plugins Does Anthropic Provide?

Anthropic released 11 open-source plugins at launch. Each targets a specific job function with pre-built skills, commands, and connectors.

Marketing Plugin

The marketing plugin handles content drafting, campaign planning, brand voice enforcement, competitor briefing, and performance reporting. Connectors include Canva, Figma, HubSpot, Amplitude, Notion, Ahrefs, SimilarWeb, and Klaviyo.

Commands include /marketing:campaign-brief, /marketing:content-calendar, and /marketing:competitor-brief. Skills encode brand voice guidelines, campaign planning methodology, and channel-specific best practices.

Sales Plugin

The sales plugin researches prospects, preps for calls, reviews pipeline, drafts outreach, and builds competitive battlecards. Connectors include Slack, HubSpot, Close, Clay, ZoomInfo, Notion, Jira, Fireflies, and Microsoft 365.

Productivity Plugin

The productivity plugin manages tasks, calendars, daily workflows, and personal context. Connectors include Slack, Notion, Asana, Linear, Jira, Monday, ClickUp, and Microsoft 365.

Data Plugin

The data plugin queries, visualizes, and interprets datasets. It writes SQL, runs statistical analysis, builds dashboards, and validates work before sharing. Connectors include Snowflake, Databricks, BigQuery, Hex, Amplitude, and Jira.

Other Plugins

Additional plugins cover customer support (ticket triage, response drafting), product management (specs, roadmaps, user research), finance (journal entries, reconciliation), legal (contract review, compliance tracking), enterprise search (cross-tool information retrieval), bio research (literature search, genomics analysis), and plugin management (creating and customizing other plugins).

All plugins are available on GitHub at anthropics/knowledge-work-plugins under Apache 2.0 license.

Action item: Browse the plugin repository at github.com/anthropics/knowledge-work-plugins. Read the marketing plugin’s skill files to understand how Anthropic structures domain expertise.

How Do You Customize Plugins for Your Team?

Default plugins provide generic starting points. Real value comes from customizing them for your specific tools, terminology, and processes.

Swap Connectors

Edit .mcp.json to point at your tool stack. If you use Pipedrive instead of HubSpot, update the CRM connector. If you use Mixpanel instead of Amplitude, change the analytics connector.

The connector file maps tool names to MCP server commands. Claude calls tools by name; the connector routes those calls to the correct MCP server.

Add Company Context

Drop your terminology, org structure, and processes into skill files. Generic skills use generic terms. Your skills should use your vocabulary.

Replace “target audience” with your actual segment names. Replace “campaign types” with your specific campaign categories. Replace “approval workflow” with your actual stakeholder chain.

Adjust Workflows

Modify skill instructions to match how your team operates. Default skills follow textbook processes. Your team has specific sequences, validation steps, and handoff points.

Add your checklist items. Insert your quality gates. Include your tool-specific gotchas (like HubSpot’s semicolon requirement for multi-select fields).

Build New Plugins

Use the cowork-plugin-management plugin or follow the file structure to create plugins for workflows Anthropic has not covered. Package your tribal knowledge into shareable, version-controlled files.

Marketing operations, RevOps, partner marketing, event marketing, and ABM workflows all make strong plugin candidates. Any repeatable process with multiple steps benefits from plugin packaging.

Action item: Fork the marketing plugin from GitHub. Edit one skill file to include your brand voice guidelines. Replace connector configuration with your actual tool stack. Upload the customized plugin to Cowork.

What Is the Decision Matrix for Claude Products?

Different tasks suit different Claude interfaces. Use this matrix to route work appropriately.

Quick Content Drafts

Use Claude.ai Web. Fast iteration with Artifacts for live preview. No setup required. Best for ad copy, social posts, email drafts, and short-form content.

Sensitive Data Analysis

Use Claude Desktop with local MCP. Data stays on your machine. No cloud upload. Best for customer lists, financial data, and competitive intelligence you do not want leaving your network.

Repeatable Automation

Use Claude Code or Cowork with plugins. Persistent workflows with sub-agent orchestration. Best for campaign launches, reporting cycles, and multi-step processes you run repeatedly.

Team Collaboration

Use Claude.ai Projects. Shared context with activity feeds and role-based permissions. Best for team knowledge bases, shared campaign contexts, and collaborative content development.

Multi-Tool Orchestration

Use Cowork with plugins. Sub-agent coordination across multiple MCP servers. Best for workflows spanning CRM, analytics, email, and project management tools simultaneously.

Enterprise Deployment

Use Claude Enterprise with organization-wide plugin deployment. Admin controls, SSO, audit logs. Best for standardizing AI workflows across departments.

Action item: Audit your last week of Claude usage. Categorize each task by this matrix. Identify tasks you handled in Chat that should move to Cowork for autonomous execution.

What Is the Getting Started Path for Marketing Teams?

Follow this four-week path to move from chat-based prompting to plugin-based automation.

Week 1: Foundation

Download Claude Desktop for macOS. Create a Project in Claude.ai called “Brand Voice” and upload your style guides, messaging frameworks, and content examples. Test content generation with explicit brand instructions.

Create your first skill folder with brand voice guidelines. Upload to Settings > Capabilities > Skills. Verify the skill triggers when you mention brand-related tasks.

Week 2: First Plugin

Identify your most repetitive workflow. Email sequences, campaign briefs, and competitive analysis make strong candidates. Use the cowork-plugin-management plugin to generate a starter plugin.

Test the plugin with 10 variations of triggering requests. Refine the skill instructions based on output quality. Add your specific templates and quality standards.

Week 3: MCP Integration

Connect your primary MCP tools. Start with Slack or your project management platform. Test data retrieval by asking Claude to “check my recent messages” or “list my open tasks.”

Create a skill that uses the MCP connection. Combine tool access with your workflow methodology. Test end-to-end execution from trigger to output.

Week 4: Cowork Workflows

Move a complex workflow into Cowork. Choose something with 5+ steps and multiple tool touches. Build the workflow as a command in your plugin.

Test autonomous execution. Provide the trigger, approve the plan, step away, return to finished work. Measure time saved versus manual execution.

Action item: Block 2 hours this week for Week 1 tasks. Download Claude Desktop, create the Brand Voice project, and build your first skill. The foundation determines everything that follows.

Final Takeaways

Claude’s product ecosystem separates into interfaces (Claude.ai, Desktop, Code) and capabilities (Skills, MCP, Cowork, Plugins). The interfaces determine how you interact; the capabilities determine what Claude executes.

Skills are pre-written instructions that eliminate repetitive explanation. One skill file replaces hundreds of context-setting messages across conversations.

MCP provides tool connections while skills provide methodology. Connecting HubSpot gives Claude access. A HubSpot skill teaches Claude your sales process.

Cowork brings autonomous execution to non-developers. Describe outcomes, approve plans, receive finished work. The conversation-to-delegation shift changes how marketing teams operate.

Plugins bundle skills, commands, connectors, and sub-agents into role-specific packages. Install once, configure for your tools, and Claude becomes a marketing specialist who knows your processes.