AI Productivity Guide · Claude Cowork · 2025
📖 Complete Reference Guide
20 Claude Cowork Concepts
Every beginner needs to know
Skills, MCPs, sub-agents, connectors — everyone throws these words around like you already know what they mean. This guide breaks down all 20 Claude Cowork concepts in plain English, from the absolute basics to the most powerful automation features, in a single definitive reference.
📅 Updated June 2025⏱ 18 min read🎯
Beginner → Advanced🧠
20 concepts covered
Claude Cowork is one of the most powerful AI productivity tools available today — but if you’re new to it, it can feel completely overwhelming. Skills, MCPs, sub-agents, connectors… this guide cuts through every term and concept, starting from zero and building up to the platform’s most sophisticated features.
Foundational Concepts (Start Here)
Concept 01 / 20
The Workspace Folder
Beginner
The workspace folder is the home base of everything Claude Cowork does on your computer. It is a real folder on your local file system — Mac or Windows — that Claude has direct access to read from and write to. Every file Claude generates lives here: HTML dashboards, Python scripts, CSV reports, Word documents, and more.
To get started, download the Claude desktop app, open Cowork mode using the toggle at the top of the interface, and either select an existing folder or create a new one (e.g. “Claude Cowork Folder”). Once linked, that folder becomes Claude’s working environment for every task you assign it.
Pro tip You can link multiple workspace folders as your projects expand — one for client work, one for personal automation, one for content creation.
Concept 02 / 20
CLAUDE.md — Your Project Brain
Beginner
CLAUDE.md is a plain markdown file that lives inside your workspace folder. Claude reads it automatically at the start of every single session — before any message you send. Think of it as a permanent, project-specific system prompt.
Use CLAUDE.md to tell Claude who you are, what rules it must always follow, what tone to use, and any critical context about the project it’s operating in. For example, a CLAUDE.md for a client agency might include the client’s brand voice, pricing tiers, team member names, and output formatting rules.
What to include in your CLAUDE.md
- 👤Who you are and what you do (creator, agency owner, developer, etc.)
- 📋Rules: “Always respond in bullet points,” “Never use passive voice”
- 🏢Business context: pricing, services, team structure, client names
- 🎯Output formats: preferred file types, design styles, dashboard templates
Key distinction CLAUDE.md is project-specific and Cowork-only. It is different from Global Instructions (see Concept 3). Each project folder can have its own unique CLAUDE.md.
Concept 03 / 20
Global Instructions — Your Permanent Identity
Beginner
While CLAUDE.md applies to one specific project, Global Instructions apply to every Claude session — Cowork, chat mode, everything. They are your universal preferences baked into the entire Claude platform.
Access them by going to your plan settings → Cowork → Global Instructions. Describe your identity, preferred communication style, and any conventions that should always apply regardless of project context.
| Feature | CLAUDE.md | Global Instructions |
|---|---|---|
| Scope | One project folder | All of Claude |
| Applies to chat mode? | No | Yes |
| Where it lives | Inside the workspace folder | Claude settings panel |
| Best for | Project-specific context | Universal identity & tone |
Concept 04 / 20
Memory — Claude Remembers What Matters
Beginner
Memory in Claude Cowork works exactly as the name implies — it lets Claude retain important information across separate conversations so it doesn’t start from zero every time. Internally, memory is stored as a collection of markdown files inside your project.
Claude builds memory automatically over time, capturing preferences, workflow decisions, and context you’ve established. You can also create memories manually — just say “save that to memory” during any conversation and Claude will persist it.
- 🎨Output preferences: “Always use Apple-style minimal design for dashboards”
- 🚫Prohibitions: “Never use these phrases in my writing”
- 📊Context: recurring clients, product details, team members
- ⚙️Workflow feedback: what worked and what didn’t in previous tasks
Read Also: How to Sell Digital Products Online in 2026: A Step-by-Step Guide for Beginners – AI Discoveries
Concept 05 / 20
The Context Window — Claude’s Working Desk
Beginner
The context window is the total amount of information Claude can actively hold in mind during a single conversation. Think of it as a physical desk — you can only fit so much on the surface at once.
As of 2025, Claude supports a 1 million token context window. Everything loaded into a session counts toward this limit: the CLAUDE.md file, memory files, your conversation history, uploaded documents, tool outputs, and system prompts.
Practical implication In very long sessions, Claude may need to compress or deprioritize older context. For complex long-running projects, use Projects (Concept 10) with targeted memory files rather than one endless conversation.
Tier 2
Core Capabilities
Concept 06 / 20
Multimodal — Claude Can See, Not Just Read
Beginner
Multimodal means Claude can process more than just text. It can analyse images, screenshots, charts, PDFs, diagrams, and data files — not just the words you type into the chat box.
When you upload a screenshot of a spreadsheet, an image of a whiteboard, or a PDF contract, Claude processes the visual information just as it processes text. The one notable exception: Claude cannot yet process video files.
- 🖼️Upload a competitor’s website screenshot → get a breakdown and improvement suggestions
- 📄Drop in a PDF contract → get a redline review with negotiation notes
- 📊Share a chart image → receive data interpretation and recommended next steps
Concept 07 / 20
Web Search — Real-Time Information Access
Beginner
Claude’s training data has a knowledge cutoff — it doesn’t inherently know about events, pricing changes, or new AI tool releases after that date. The web search tool solves this by letting Claude query the live internet in real time.
When Claude encounters a question requiring current information — today’s news, the latest model pricing, a recently published article — it fires off a web search automatically and incorporates the results into its response.
Concept 08 / 20
Extended Thinking — Deliberate, Deep Reasoning
Intermediate
Extended thinking allows Claude to work through unusually complex problems with greater deliberation before producing a final answer. Rather than replying immediately, Claude reasons through the problem in an internal monologue — weighing options, checking for contradictions, and exploring multiple paths.
The tradeoff is speed: extended thinking takes longer but produces noticeably better outcomes for hard analytical tasks, complex decision trees, multi-step reasoning, and ambiguous strategy questions.
When to use it Use extended thinking for business strategy, complex code architecture, contract analysis, or any task where a rushed answer would cost you more time later than the thinking delay costs you now.
Concept 09 / 20
Artifacts — Outputs You Can Actually Use
Intermediate
An artifact is any structured output Claude produces beyond a text response — a rendered HTML dashboard, an interactive chart, a slide deck, a formatted report, a working React component. Artifacts are not just code snippets; they are live, viewable outputs inside the Cowork interface.
You can open any artifact in your browser, download it, or embed it in other workflows. Claude generates artifacts automatically whenever a task calls for visual or interactive output.
- 📊HTML competitive intelligence dashboards that auto-update
- 📑Interactive slide presentations built entirely in Claude
- 📋Dynamic tables aggregating data from multiple connected sources
- 🖥️Functional web app prototypes generated from a text description
Concept 10 / 20
Projects — Separate Brains for Every Context
Intermediate
Projects are organizational containers inside Claude Cowork. Each project has its own memory, its own CLAUDE.md instructions, its own conversation history, its own scheduled tasks, and its own file library. They are separate brains that prevent unrelated work from bleeding together.
A content creator might run separate projects for YouTube, Instagram, a client agency, and their personal brand. Each project understands its own context completely without the others interfering.
Best practice Create a project for every major context in your life or business where you’ll run recurring tasks. Projects are where Claude Cowork shifts from a chat tool to a genuine AI operating system.
Tier 3
Power User Features
Concept 11 / 20
The Bash Tool — Code That Runs Itself
Intermediate
The bash tool is how Claude Cowork executes code without you ever opening a terminal. When a task requires computation — renaming files in bulk, resizing images, processing CSV data, running a Python script — Claude writes the code and runs it automatically using the bash tool.
You don’t need to understand the code or copy it anywhere. Claude handles the full loop: write, execute, check output, iterate if needed.
- 🗂️Rename 200 files in a folder according to a specific naming convention
- 🖼️Resize and compress a batch of product images to web specifications
- 📈Aggregate and analyse CSV sales data across multiple files
- 🔍Scan a codebase for deprecated function calls and generate a report
Concept 12 / 20
Skills — One Command, Entire Workflows
Intermediate
Skills are the single most powerful feature in Claude Cowork. A skill is a saved markdown file that encodes a complete, reusable workflow — including every step, every output format, every integration, and every preference.
Once a skill is created, you trigger the entire workflow with a single phrase. A “Morning Briefing” skill, for example, might automatically: scan your Google Calendar, check your email, pull the top trending AI news, write your three daily priorities, generate a formatted HTML dashboard, and post a summary to Slack — all from one command.
Examples of high-value skills
- 🌅Morning Briefing — Calendar + email + news + priorities → HTML dashboard + Slack post
- 📱Instagram Carousel Generator — Topic prompt → branded carousel slides → auto-published to Instagram
- 📄Contract Reviewer — Upload PDF → redlined review with negotiation priorities
- 🧾Invoice Generator — Client data → formatted invoice PDF → emailed automatically
- 🎙️Proposal Generator — Brief notes → full client proposal document
Concept 13 / 20
Slash Commands — The Trigger Layer
Intermediate
Slash commands are the keywords that activate skills. When you type /morning-briefing, Claude looks up the skill file with that name and immediately executes the entire workflow it encodes.
They sound technical but the concept is simple: a slash command is just a shortcut. The slash prefix distinguishes workflow triggers from regular conversation, so Claude knows you want to run a predefined process rather than chat.
Concept 14 / 20
Plugins — Shareable Bundles of Skills
Intermediate
If a skill is a single workflow, a plugin is a packaged collection of multiple skills bundled into a shareable zip file. Plugins let you export your entire skill library, share it with teammates, or install skill packs created by others — all in one upload.
Anthropic provides official plugins for specific industries (Legal, Design, Marketing, etc.) that you can browse directly inside Cowork. Installing a plugin instantly adds every skill it contains to your Cowork environment.
| Feature | Skill | Plugin |
|---|---|---|
| What it contains | One workflow | Many workflows (skills) |
| Triggered by | Slash command | Installing the bundle |
| Shareable? | Manual file share | Yes — single zip file |
| Best for | Individual custom automation | Full workflow suites & team sharing |
Tier 4
Advanced Automation & Agent Features
Concept 15 / 20
Connectors — Linking Your Entire App Stack
Advanced
Connectors are the integration layer that allows Claude Cowork to take real actions inside your existing tools — reading emails, adding calendar events, posting to Slack, querying a CRM, updating a spreadsheet.
Anthropic provides native connectors for popular platforms (Google Calendar, Gmail, Slack, QuickBooks, Gusto, and more) accessible directly from the Customize → Connectors menu.
For platforms not yet natively supported, you can connect via a custom MCP server (Model Context Protocol) using tools like Zapier MCP or Composio. This extends Claude’s reach to virtually any application with an API.
Integration hierarchy For any given app, prefer connectors first (fastest, most reliable), browser automation second, computer use third. Connectors are always the most stable and efficient pathway.
Concept 16 / 20
Claude in Chrome — Browser-Level Automation
Advanced
The Claude Chrome extension gives Claude the ability to operate inside your web browser — navigating pages, clicking elements, filling out forms, extracting content, and performing actions on any website you’re logged into.
This is particularly valuable when a connector doesn’t exist for a platform you use. If Claude can’t connect to a platform’s API, it can often accomplish the same goal by interacting with the platform’s web interface directly — the same way you would as a human user.
- 📋Extract the latest posts from a community platform with no API access
- ✍️Fill out multi-step web forms on your behalf
- 🔗Scrape structured data from websites for research and reporting
- 🛒Book flights, make reservations, or complete checkout flows
Concept 17 / 20
Computer Use — Full Desktop Control
Advanced
Computer use extends Claude’s autonomy beyond the browser to your entire desktop. Claude can open native applications, navigate your file system, drag and drop files, interact with desktop software like video editors or design tools — anything a human could do with a keyboard and mouse.
Grant Claude temporary access to your screen and it will execute tasks hands-free: finding a video file on your desktop, opening a video editor, and uploading the clip to the timeline — all without you touching the keyboard.
Current limitation Computer use is powerful but slower than connectors or browser automation. It is best reserved for tasks where no API or browser path exists. Speed and reliability will improve over time.
Concept 18 / 20
Scheduled Tasks — Autonomous Recurring Automation
Advanced
Scheduled tasks allow Claude to run skills automatically at defined times — daily, hourly, weekly — without any manual trigger from you. They are the bridge between Claude Cowork and traditional automation platforms like n8n or Zapier.
Configure a scheduled task either by setting up a new task directly in the Scheduled panel, or simply by telling Claude during a conversation: “Run this every morning at 9am” — and it will convert the current session into a recurring scheduled automation.
- ☀️Daily briefing at 7am — Calendar + email + news → dashboard posted to Slack
- 🌙Evening wrap-up — Summary of everything Cowork did today → HTML report
- 📡Community pulse check — Scan for new posts and comments needing replies → digest email
- 📉Competitor monitoring — Pull new competitor content and flag notable changes weekly
Concept 19 / 20
Dispatch Mode — Remote Control from Your Phone
Advanced
Dispatch mode is a mobile-first interface that lets you control Claude Cowork — and by extension your computer — from your smartphone. While you are away from your desk, you can instruct Claude to perform desktop tasks and they will execute on your home or office machine.
A unified conversation view in Cowork on desktop mirrors exactly what you see in Dispatch on your phone, keeping all your sessions in sync. Dispatch also supports triggering any skill you have installed — making your full automation library available from anywhere.
Example use case You’re at the gym. You text Dispatch: “Upload this week’s podcast recording to Google Drive and send the link to my editor.” Your computer handles it autonomously.
Concept 20 / 20
Sub-Agents — Parallel AI Workers at Scale
Power Feature
Sub-agents are the most powerful concept in the entire Claude Cowork architecture. They allow a single complex task to be broken into parallel workstreams, each handled by a dedicated specialised agent simultaneously.
Instead of Claude executing steps sequentially — research, then write, then design — sub-agents work in parallel. One agent handles market research while another pulls financial data while a third writes the narrative while a fourth builds the slide deck. The orchestrating agent then synthesises everything into a final output.
Sub-agent workflow example: Investor pitch preparation
- 🔵Agent 1 — Conducts market sizing and competitive landscape research
- 🟠Agent 2 — Pulls and formats financial projections from connected spreadsheets
- 🟢Agent 3 — Writes the narrative arc and key talking points
- 🟣Agent 4 — Builds the slide deck artifact with branded design
- ⚪Orchestrator — Assembles all outputs into a unified, investor-ready presentation
Why this matters Sub-agents compress hours of sequential work into minutes of parallel execution — and the quality improves because each agent specialises in its assigned domain rather than context-switching.
All 20 Concepts at a Glance
01Workspace FolderClaude’s local file system home base
02CLAUDE.mdProject-specific instructions loaded every session
03Global InstructionsUniversal preferences across all of Claude
04MemoryCross-session context that builds over time
05Context WindowActive working memory — currently 1M tokens
06MultimodalImages, PDFs, charts — not just text
07Web SearchReal-time internet access for current data
08Extended ThinkingDeep deliberation for complex problems
09ArtifactsRendered visual and interactive outputs
10ProjectsIsolated brains per business context
11Bash ToolCode that writes and runs itself
12SkillsOne-command complete workflow execution
13Slash CommandsKeyword triggers that activate skills
14PluginsBundled, shareable skill collections
15ConnectorsNative integrations with external apps
16Chrome ExtensionBrowser-level web automation
17Computer UseFull desktop application control
18Scheduled TasksAutomated recurring workflow execution
19Dispatch ModeRemote desktop control from mobile
20Sub-AgentsParallel AI workers for complex tasks
Frequently Asked Questions
What is Claude Cowork and how is it different from regular Claude chat?
Claude Cowork is a dedicated agentic workspace within the Claude desktop app. Unlike the regular chat interface — which is conversational and single-session — Cowork gives Claude access to your local file system, lets it run code, connect to external applications, maintain persistent memory, and execute scheduled automations autonomously. It is Claude operating as an AI employee, not just a chatbot. Do I need to know how to code to use Claude Cowork?
No. Claude Cowork is designed for non-technical users. The bash tool writes and runs code on your behalf. Skills are created in plain markdown — structured writing, not programming. Connectors are set up via a simple GUI. You describe what you want in plain English and Claude handles the technical execution. What is the difference between Claude Cowork Skills and Plugins?
A skill is a single saved workflow triggered by one slash command. A plugin is a zip file containing many skills bundled together. Plugins are how you share, distribute, or install entire workflow suites at once. Anthropic also publishes official plugins by industry — Legal, Marketing, Design — that you can install from the Cowork interface directly. What are MCP servers and do I need them?
MCP (Model Context Protocol) servers are the underlying protocol that powers connectors. When you connect Claude to Gmail or Google Calendar natively, Anthropic is running an MCP server in the background. You only need to set up a custom MCP server (using tools like Zapier MCP or Composio) if you want to connect to an application that isn’t yet available as a native Anthropic connector. Most users won’t need to configure this manually. Are scheduled tasks the same as automation platforms like n8n or Zapier?
They serve a similar purpose — running automated workflows on a schedule — but Claude’s scheduled tasks are more flexible for AI-driven tasks because the agent itself can reason, adapt, and make decisions mid-execution. Traditional automation platforms excel at rigid, deterministic workflows; Claude’s scheduled tasks excel when the task requires understanding context, writing prose, or making judgment calls with incomplete information. How many sub-agents can run simultaneously in Claude Cowork?
Anthropic has not published a fixed hard limit. In practice, sub-agent parallelism scales with the complexity of the task and the skills you have built. Each sub-agent draws on its own assigned skill and context, coordinated by an orchestrator agent that assembles the final output. Is Claude Cowork available on Windows and Mac?
Yes. The Claude desktop app — which includes Cowork — is available for both macOS and Windows. Download it from anthropic.com/claude.
What to Build First
If you have made it through all 20 concepts, you now have a genuine mental model of what Claude Cowork can do — from basic file access all the way to parallel sub-agent pipelines. The concepts stack on each other deliberately: you cannot build powerful skills without first understanding the workspace and CLAUDE.md; you cannot run scheduled automations without first understanding connectors.
The best starting point for most people is this sequence: set up your workspace folder and write a CLAUDE.md → connect Google Calendar and Gmail as connectors → build your first skill (a morning briefing is the classic starting point) → schedule it to run automatically every morning. That single workflow will give you an immediate, tangible return on the time you spent learning these 20 concepts.
Everything else — sub-agents, computer use, dispatch mode, plugins — is an extension of that foundation. Build incrementally, and the complexity never becomes overwhelming.
Claude Cowork Complete Guide · Built for clarity, cited for accuracy · 2026
Leave a Reply