🧠 How to Use Claude Cowork: Workflows, Examples, and Setup Guide

TL;DR — Too Long Didn’t Read

Skip the guide and do this:

• Cowork is an autonomous task-execution mode — not a better chat. It reads and writes files, runs multi-step workflows, and can operate while you step away.

• The biggest mistake people make is prompting Cowork like they prompt Chat. Cowork needs scope, constraints, and clear deliverables — not open-ended questions.

• No memory between sessions is the default. You have to build memory yourself using context files.

• Global Instructions and a dedicated project folder are the two setup moves that matter most before you do anything else.

• Plugins bundle skills + connectors + sub-agents into one package. They're the fastest way to make Cowork feel tailored to your work.

• Scheduled tasks are powerful but require your computer to be awake and Claude Desktop to be open to run.

• Cowork is a research preview. Don't use it for regulated workloads, and don't give it access to sensitive files you aren't comfortable with it reading.

1. What Is Claude Cowork?

Cowork brings the agentic capabilities of Claude Code to non-technical users through Claude Desktop. The simplest way to understand it: standard Claude Chat responds to what you paste in the moment. Cowork takes a goal, plans a sequence of steps, reads from your actual files, creates new ones, and keeps working until there's a deliverable you can open.

The output is usually something tangible — a formatted report, an organized folder, a filled spreadsheet, a drafted document — not text in a chat window you then have to act on yourself.

A few things that make this different from Chat in practice:

• Direct file access. You grant Cowork permission to a folder on your computer. It reads across all files in that folder without you uploading anything. This matters when your work lives across dozens of documents.

• Sub-agent coordination. For complex tasks, Cowork breaks the work into parallel workstreams and runs them simultaneously. Processing 10 files in parallel instead of one at a time can cut a 30-minute task down to around 4 minutes.

• Long-running execution. Cowork doesn't time out the way a chat session would. You can describe a task, walk away, and come back to finished work.

• Scheduled automation. Cowork can run tasks on a repeating schedule — daily briefings, weekly summaries, recurring file organization — without you having to kick it off each time.

What Cowork is not: it isn't smarter than Chat, and it doesn't have memory between sessions by default. Every new session starts fresh unless you build context into files it can read at the start.

Claude Cowork vs Chat: When Should You Use Each?

The rule of thumb is simple: if your task requires Claude to make decisions, produce files, or work across multiple documents without you driving it step by step — use Cowork. If you want a back-and-forth, an answer, a draft you'll refine, or something you haven't fully scoped yet — use Chat.

Bonus: We put together a detailed guide on when to use each tool and why it works.

2. How to Set Up Claude Cowork (Step-by-Step)

Most people open Cowork and immediately try a task. That works, but you'll hit the same friction every session: re-explaining your context, re-stating your preferences, re-scoping your work. Two setup steps eliminate most of that friction.

Step 1: Write Your Global Instructions

Navigate to Settings > Cowork in Claude Desktop and add Global Instructions. These apply to every Cowork session. Think of them as your standing brief — the things you'd tell a new assistant on day one.

What to include:

• Who you are and what your work involves

• Output format preferences (length, structure, tone)

• Things you always want Claude to ask about before taking action

• Things you never want it to do (delete files without confirming, make assumptions about scope)

Keep this focused. The more specific and concise your instructions, the more consistently Cowork follows them. Dense paragraphs get skimmed; structured bullets get used.

Step 2: Create a Dedicated Project Folder with Context Files

Don't grant Cowork access to your entire Documents folder on day one. Create a dedicated working folder and give Claude access only to that. This limits exposure and makes behavior more predictable.

Inside that folder, create a subfolder called _context (or similar) and populate it with a few lightweight files:

about-me.md — Your role, your team, your industry, current priorities, and what "good work" looks like for you. A few paragraphs is enough.

voice-and-style.md — How you want things written. Tone preferences, formatting rules, words or phrases to avoid, examples of writing you like. This is particularly useful if Cowork is generating anything you'll send or publish.

memory.md — Since Cowork has no memory between sessions, use this file to log ongoing context: active projects, decisions already made, things in progress. Update it at the end of sessions or ask Claude to update it for you before the session ends.

current-projects.md — A short list of active workstreams with one-line summaries. Helps Cowork orient itself without you having to re-brief every time.

These files compound. The more you maintain them, the less re-explaining you do. Many users find output quality improves noticeably within a few weeks just from keeping these files current.

Step 3: Configure Folder Instructions

Beyond Global Instructions, you can add folder-level instructions when you select a local folder. These are project-specific and override or supplement your global settings for that context. If you work on multiple different project types, folder instructions let you keep the context tight and relevant rather than loading everything into one global file.

3. How to Write Effective Claude Cowork Tasks

The single biggest predictor of a useful Cowork session is how you frame the task upfront. Vague prompts cause drift. When Claude has to make decisions about file operations and outputs, ambiguity leads to outputs you didn't ask for, or worse — actions that are hard to undo.

Think of it less like prompting a chatbot and more like writing a job spec for a colleague. Include:

• Scope — What files or data is in play? What's out of scope?

• Constraints — What should Claude not do? Should it ask before deleting, overwriting, or moving anything?

• Deliverables — What does done look like? A specific file? A folder structure? A formatted document?

• Format — How should outputs be structured? Bullet list? Table? Specific headings?

Here's the difference in practice:

❌ "Summarize my project notes"

✅ "Read all .md files in the /projects/q2-planning folder. Create a single document called q2-summary.md that pulls out: decisions already made, open questions, and next actions. Use those three as H2 headers. Don't delete or modify the source files."

The second version takes 30 seconds longer to write and produces dramatically more reliable output.

A few other patterns that help:

Ask for source tracing. If Cowork is synthesizing across files, ask it to cite the filename next to each key claim. This makes outputs auditable and easier to verify.

Build in a checkpoint. For long tasks, ask Claude to pause and confirm before taking any irreversible action — moving files, sending messages, modifying originals.

Be explicit about uncertainty. Tell Claude to flag anything it's unsure about rather than making assumptions. "If the scope is unclear, ask me before proceeding" is a line worth including in most task prompts.

4. Claude Cowork Plugins, Skills, and Connectors Explained

These three are related but distinct, and understanding the difference helps you build a setup that actually fits how you work.

Connectors

A connector is a single integration between Claude and one external tool. Gmail, Slack, Notion, Google Drive, Linear, and others are available in the Connectors Directory.

Skills

Skills are sets of instructions that tell Claude how to behave for specific tasks — output formats, workflows, domain-specific behavior. They can be attached to plugins, or you can create your own. When a skill is relevant to what you're working on, Claude uses it automatically. You can also trigger skills manually by typing / in the task interface.

Plugins

Plugins bundle skills, connectors, and sub-agents into one package. Instead of setting each piece up individually, a plugin gives you a pre-configured setup for a specific role or workflow — sales, finance, legal, marketing, HR, engineering, and others are available in the plugin library.

Bonus: We put together a guide explaining these features and how to use them effectively.

5. How to Use Scheduled Tasks in Claude Cowork

Scheduled tasks are one of the most underused features in Cowork. You describe a recurring job once, and Cowork runs it automatically — daily briefings, weekly report compilation, regular file cleanup, whatever repeats in your work.

Setting One Up

Method 1 (quick): Type /schedule in any task, describe what you want, answer Claude's clarifying questions, confirm.

Method 2 (structured): Click Scheduled in the sidebar > + New task > fill in: task name, description, the full prompt, frequency (hourly / daily / weekly / weekdays / manual), and optionally a specific folder or model.

The Critical Limitation

Scheduled tasks only run while your computer is awake and Claude Desktop is open. If your computer is asleep or the app is closed when a task is scheduled to run, Cowork skips that execution — then runs it automatically once the system is active again.

This makes it less useful for anything where timing matters precisely (like a 9am daily brief if you don't open your laptop until 10). Plan your schedules with this in mind, or use the "manual" frequency for tasks you want on demand rather than automatic.

Good use cases for scheduled tasks:

• Daily summary of new emails by category

• Weekly compilation of data from a recurring spreadsheet

• Morning briefing pulling notes from the previous day's files

• Regular folder cleanup by file type or date

What to avoid scheduling:

• Tasks that involve sensitive files you don't want running automatically

• Anything that sends messages or takes external actions unless you've thoroughly tested it first

• Tasks that touch original files without backups — schedule tasks that read and create, not overwrite

Review the run history for scheduled tasks regularly. Things drift over time, especially if your folder structure or file formats change.

6. Safety: What to Watch and What to Avoid

Cowork is a research preview. Agent safety is still in development, and Anthropic says as much on the product page. This doesn't mean don't use it — it means use it with appropriate awareness of what it can and can't do on your behalf.

File Access

Restrict the folder you grant access to. A dedicated project folder is much safer than pointing Cowork at your entire Documents directory. If a task goes sideways, a narrow scope limits the blast radius.

Avoid granting access to folders containing financial documents, credentials, HR records, or anything you'd consider sensitive. Even if you trust the output, the data moves through a system that's still in preview.

Prompt Injection

When Cowork reads web content or documents from untrusted sources, those files can contain embedded instructions meant to manipulate what Claude does next. Limit Cowork's browsing to sites you trust. Be especially cautious with tasks that ask Claude to read external documents or emails and then take action based on their content.

Connectors and Permissions

Set connector permissions to "Ask" before allowing Claude to take actions you'd want to review — especially for anything involving sending messages, creating content that others will see, or modifying records in external systems. Start with Allow only for read-only actions, and expand permissions only after you've seen how the connector behaves.

Regulated Workloads

Cowork activity is not captured in Audit Logs, the Compliance API, or data exports. Do not use Cowork for work that requires auditability or regulatory compliance.

You're Still Accountable

Everything Cowork does happens under your account and on your behalf. If it publishes content, sends a message, or modifies a file — that's on you. This isn't a reason to avoid Cowork, but it is a reason to build in checkpoints for anything you haven't thoroughly tested.

7. Common Mistakes and How to Fix Them

Treating Cowork like Chat.

Cowork needs a job spec, not a question. Reframe your prompts around scope, deliverables, and constraints.

Starting with no context setup.

If you skip Global Instructions and context files, you're re-explaining yourself every session. Take 30 minutes to set these up once and update them as your work changes.

Granting access to too much.

Give Claude access to a dedicated folder, not your whole computer. If you need to expand scope, do it deliberately per session.

Running everything through Cowork.

Simple requests should go to Chat. Using Cowork for things Chat can handle burns through your usage allocation faster and doesn't produce better results.

Vague prompts on long tasks.

The longer the task, the more important the spec. A vague 2-minute task is recoverable. A vague 30-minute task might produce something that needs to be redone entirely.

Never reviewing scheduled task outputs.

Scheduled tasks drift if your file structure or expectations change. Check outputs periodically, not just when something breaks.

Installing many plugins at once.

Add one plugin, customize it, use it for a week, then decide if you need another. Stacking plugins before you understand any of them creates unpredictable behavior and makes it harder to tell what's influencing Claude's actions.

8. FAQs

Does Cowork have memory between sessions?

No. Each session starts blank. To get continuity, maintain a memory.md file in your working folder and either update it yourself or ask Claude to update it before closing a session.

How much does Cowork use vs. Chat?

Significantly more. Complex multi-step tasks are compute-intensive. Cowork is built for tasks that genuinely benefit from extended execution — not as a replacement for Chat on everyday requests.

Can I use Cowork on my phone?

Pro and Max users can assign tasks via mobile, but the actual execution still happens on your desktop. Your computer must be awake and Claude Desktop must be open.

Can I share a Cowork session with a teammate?

No. Sessions can't be shared. On Team or Enterprise plans, admins can distribute plugins so everyone has the same starting configuration, but sessions themselves are per-user.

What happens if I close Claude Desktop mid-task?

The session ends. If the task was partially complete, outputs generated up to that point may be saved, but Claude won't continue. For scheduled tasks, it will re-run the next time the app is open.

Can Cowork browse the internet?

Yes, when used with the Claude in Chrome extension. Limit browsing to sites you trust — web content is the primary vector for prompt injection.

Are Cowork activities logged for compliance purposes?

No. Cowork is not suitable for regulated workloads. Activity is not captured in Audit Logs, the Compliance API, or data exports.

Can I build my own plugins?

Yes. Cowork includes a Plugin Create tool, and Anthropic's full plugin library is on GitHub if you want to fork and customize existing ones.

9. Next Steps

If you've read this far and haven't done the setup work yet, that's the actual next step:

1. Write your Global Instructions in Settings > Cowork

2. Create a _context folder with an about-me.md, voice-and-style.md, and a memory.md file

Once those exist and you've run a few tasks that produce reliable output, then consider:

• Installing one plugin relevant to your most common Cowork use case and customizing it

• Adding one or two connectors with conservative permissions ("Ask" for any write actions)

• Setting up one scheduled task for something that currently takes you manual time each week

The users who get the most out of Cowork aren't the ones who've installed the most plugins or connected the most tools. They're the ones who've spent time giving Claude the context it needs to do their specific work — and who've built habits around reviewing and refining that context over time.

That's the actual superpower here. Not the features — the setup behind them.