The Complete Guide to Claude Cowork

You’ve heard people mention Cowork alongside Claude’s regular chat, but it’s not immediately obvious what it adds. Is it simply a more advanced chat window? A separate app? Something designed mainly for developers?

Cowork is a different way of working with Claude.

Instead of asking a question and reading an answer, you can hand over a larger piece of work — gathering information, working across files and tools, completing several steps, and producing a finished result.

This guide covers what Claude Cowork is, when to use it instead of regular Chat, how to set it up, and how its main features fit into real business workflows.

The Core Idea: Delegating, Not Just Chatting

Regular Claude Chat is useful when you want to think something through with AI.

You might use it to:

• draft something
• brainstorm ideas
• ask questions
• analyse information
• refine a piece of work through several messages

It works naturally as a conversation: you ask, Claude responds, and you continue from there.

Cowork is designed for a different shape of task.

Instead of working through every stage yourself, you describe the outcome you need and give Claude access to the relevant files, tools, or information.

Claude can then plan the work, complete several steps, use the resources it needs, and produce the finished deliverable.

The distinction becomes clearer with an example.

A consultant preparing a client report might normally:

1. open several source documents
2. find the relevant information
3. compare the figures
4. identify the important changes
5. write the report
6. format the finished document

In Cowork, that can become one delegated task.

You provide the source material, explain what the final report needs to achieve, and let Claude work through the steps.

A useful rule of thumb is this:

Use Chat when you want to work something through with Claude. Use Cowork when you want to hand Claude a piece of work to complete.

Chat vs. Cowork vs. Code

Claude offers different environments for different kinds of work.

Mode	Best for	Example
Chat	Turn-by-turn thinking, questions, drafts, brainstorming, analysis, and refinement	“What are three ways I could restructure this pricing page?”
Cowork	Multi-step work across files, tools, and information that ends in a finished result	“Pull this quarter’s figures from these spreadsheets and build the board presentation.”
Code	Working with software projects, codebases, tests, files, and development workflows	“Refactor this authentication flow and run the tests.”

For many business owners, marketers, consultants, and other professionals, Chat and Cowork will be the most relevant.

Chat helps you think through the work.

Cowork helps you complete it.

The Three Patterns of Work That Suit Cowork

Not every task needs Cowork. Many simple requests are faster in regular Chat.

Cowork becomes particularly useful when a task has one or more of these characteristics.

The Task Has Several Steps

If a job requires gathering information, comparing it, making decisions about what matters, and producing something at the end, Cowork can handle that sequence as one task.

For example:

Read through the last two weeks of customer support tickets, group them by issue type, identify the three most common problems, and pull three representative customer comments for each.

That is more than a single question. It is a small workflow.

The Task Depends on Real Files

Cowork is useful when the work is built around documents, spreadsheets, PDFs, presentations, or other files.

For example:

Update our Q3 client proposal using the latest case studies and current pricing in this folder, then save the finished proposal as a new file.

Instead of manually copying pieces of every source into Chat, Claude can work with the relevant files directly.

The Task Spans Several Tools

Many business tasks don’t live in one place.

The information might be spread across email, Slack, a calendar, Google Drive, a CRM, and local documents.

When Claude has access to the appropriate tools, Cowork can bring that information together as part of one workflow.

For example:

Check what the team said about the Meridian account in Slack this week, review the relevant calendar meetings, and draft a short client status update based on both.

When the task requires several steps, several files, or several tools, it is a strong candidate for Cowork.

Where You Can Use Cowork

Cowork began as a desktop-focused experience, but Anthropic has expanded it across Claude.

Depending on your plan and the features currently available to your account, Cowork can be used through:

• Claude Desktop
• Claude on the web
• Claude mobile

Some capabilities still depend on the desktop app.

For example, if Claude needs to work with local files or use applications running on your computer, Claude Desktop needs to provide that connection.

Cloud-based Cowork sessions can continue working remotely without your computer remaining on, while tasks that depend specifically on local files, browser access, or desktop applications may require Claude Desktop to be available.

Because these features are developing quickly, check Anthropic’s current plan and platform information if a particular capability is important to your workflow.

Setting Up Cowork

You can begin with a fairly simple setup.

The important part is not connecting everything you own. It is giving Claude access to only what the task needs.

Connect a Working Folder When You Need Local Files

On desktop, you can give Cowork access to a specific folder on your computer.

Claude can work with files inside the folders you allow it to access and can create or modify files there when the task requires it.

For example, a marketing agency working on a campaign for one client could connect that client’s project folder.

That is much better than providing access to an entire Documents directory containing unrelated work.

Think of the folder as the working boundary for the task.

A well-scoped folder might contain:

• the current brief
• brand guidelines
• previous reports
• source spreadsheets
• approved images
• a template for the finished deliverable

You give Claude enough information to do the job without exposing unrelated files.

Add the Connectors You Need

Connectors allow Claude to work with information stored in other services.

Depending on what is supported for your account, these can include services such as:

• Google Drive
• Gmail
• Google Calendar
• Slack
• Microsoft 365
• project-management platforms
• CRM systems
• other business tools

Once a Connector is available, you can refer to the service naturally in your task.

For example:

Find the customer follow-up from last quarter in Gmail.

or:

Check the latest launch discussion in Slack.

The exact actions Claude can perform depend on the Connector.

Some primarily allow Claude to search or read information. Others may support actions such as creating, updating, or sending.

Check the permissions and capabilities of each connection rather than assuming all Connectors behave the same way.

Understand Permissions Before You Delegate

Cowork can do more than produce text, which makes permissions more important than they are in a standard conversation.

Depending on the task and your settings, Claude may be able to:

• read files
• create files
• edit files
• access connected services
• browse websites
• interact with applications
• perform actions in external tools

Cowork includes approval controls so you can decide how much independence Claude has.

For a new workflow, the safer approach is to keep approval requirements in place while you learn how Claude handles the task.

The more consequential the action, the more human oversight it deserves.

Drafting a summary is different from sending an email.

Reading a spreadsheet is different from changing financial data.

Creating a working document is different from permanently deleting one.

Treat Cowork the way you would treat delegation to another person: the level of supervision should match the consequences of getting something wrong.

Writing Your First Cowork Prompt

A good Cowork prompt should answer three questions:

1. What should Claude produce?
2. What should it use?
3. What matters about how the work is done?

Compare these two requests.

Weak:

Can you help me put together a competitive analysis?

Stronger:

Write a four-page memo comparing our three biggest competitors’ Q3 pricing changes. Use the analyst-call PDFs in the Q3 Competitive Review folder and follow the format of last quarter’s memo. This is for the leadership offsite, so lead with a recommendation on our new pricing tier and flag anything you can’t verify.

The stronger version identifies:

• the deliverable — a four-page memo
• the sources — the analyst-call PDFs
• the reference format — last quarter’s memo
• the audience — leadership
• the business goal — informing a pricing decision
• an important constraint — flag unverifiable information

That gives Claude much less room to make the wrong assumptions.

Cowork may still ask questions before beginning if something important is unclear.

That is useful. A clarification at the start is usually cheaper than completing the wrong task.

What Happens While Claude Works

Cowork does not mean handing over a task and ignoring everything until the end.

You can remain involved while Claude works.

Claude May Ask Clarifying Questions

If the task contains an important ambiguity, Claude may ask for more information.

For example:

• Which version of the file should it use?
• Which date range applies?
• Should the final output be Word or PowerPoint?
• Which template should it follow?

Answer these questions specifically.

They help establish the boundaries of the task before more work is completed.

Review the Plan

For larger tasks, Claude may break the work into steps.

Look at the overall approach.

Ask yourself:

• Is it using the right sources?
• Is anything unnecessary?
• Has it misunderstood the objective?
• Is it about to access something it doesn’t need?

You don’t need to inspect every technical action. Focus on whether the overall direction makes sense.

Steer the Work When Necessary

If Claude starts moving in the wrong direction, correct it.

You might say:

Use the final Q3 report, not the draft.

or:

Don’t rewrite the pricing section. Only update the figures.

or:

The audience is the client, not our internal team. Adjust the tone accordingly.

Correcting the course early is usually better than waiting for a completed deliverable and starting again.

Review the Result Like a Colleague’s Draft

A finished file is still something you should review.

Treat it as you would work prepared by a capable colleague.

Check:

Did it meet the objective?
Make sure Claude solved the problem you intended rather than a slightly different one.

Are the facts correct?
Check important names, dates, figures, quotes, calculations, and claims against the source material.

Can you trace important information back to a source?
Be cautious about any specific detail you cannot verify.

Does the output work for its intended audience?
Something can be factually correct and still be too long, too technical, too formal, or poorly structured.

If most of the result is good, ask Claude to revise the existing work rather than recreating everything from the beginning.

Scheduled Tasks

Some Cowork tasks do not need to be initiated manually every time.

Scheduled tasks allow you to define a process once and have Claude run it automatically on a recurring schedule or on demand.

Examples include:

• a weekday morning briefing
• a Friday summary of what shipped that week
• a weekly competitor update
• a monthly metrics report
• recurring research
• regular file organisation

A useful way to set one up is to run the task manually first.

Make sure:

• the sources are correct
• the output is useful
• the scope is clear
• the prompt produces the result you expect

Then turn that proven workflow into a scheduled task.

Scheduled Cowork tasks can run remotely, so cloud-based tasks do not require your computer to remain awake.

However, tasks that specifically need access to local folders or applications on your computer still depend on the relevant local access being available.

That distinction matters when deciding what to automate.

Dispatch: Starting Desktop Work From Your Phone

Dispatch is designed for situations where you are away from your computer but want Claude to work with resources on it.

You can assign a task through the Claude mobile app and have Claude work through the desktop environment using the files, connectors, plugins, and applications you have made available.

For example, from your phone you could ask Claude to:

• analyse a spreadsheet stored on your computer
• search recent Slack and email activity
• prepare a briefing document
• organise files in a particular folder
• build a presentation from material in your working folders

Because Dispatch works through your desktop environment, your computer must be awake, connected, and running Claude Desktop while the task is being carried out.

That makes Dispatch different from a remote Cowork session or a cloud-based scheduled task.

Use Dispatch when the task needs your actual desktop resources.

Use a remote Cowork session when the work can run using cloud-accessible tools and information without depending on your computer being awake.

Global Instructions

If you use Cowork regularly, you will probably notice yourself repeating certain instructions.

Perhaps you always say:

• keep reports concise
• lead with the recommendation
• avoid jargon
• use British English
• structure documents with clear headings

Global instructions let you define standing preferences that apply across Cowork sessions.

For example:

I’m an independent financial consultant working primarily with small businesses. Keep briefs concise, lead with the recommendation, and use plain language rather than unnecessary jargon.

You do not need to create a perfect set of instructions on day one.

Start using Cowork first.

When you notice yourself giving Claude the same correction repeatedly, consider adding it to your standing instructions.

Projects: Keeping an Ongoing Stream of Work Together

Projects help organise work that belongs to the same ongoing area.

You might create a Project for:

• one client
• a product launch
• a research programme
• a recurring business process
• a department
• a long-running deliverable

A Project can provide its own:

• instructions
• context
• files or linked resources
• scheduled tasks
• memory related to work within that Project

For example, a marketing agency could create a Project for a long-term client and include:

• the client’s brand guidance
• campaign history
• product information
• relevant working files
• Project-specific instructions
• recurring reporting tasks

Claude can then use that Project context when working on future tasks in the same workspace.

Project memory can also help Claude build on previous work within that Project.

However, don’t assume that every useful detail from every task will automatically become permanent Project knowledge. If something is important enough that Claude should reliably reference it in future work, add it deliberately to the relevant Project instructions or knowledge.

Projects are available across Cowork surfaces, although Projects that depend directly on folders stored on your computer require desktop access for that local material.

Skills: Teaching Claude a Repeatable Process

A Skill is a reusable set of instructions and resources that teaches Claude how to perform a particular type of task.

Think of it as a playbook.

Instead of repeatedly telling Claude:

Use these sections, follow this structure, apply this calculation, include these checks, and format the result this way.

you can create a Skill containing that process.

A Skill can be simple.

For example, it might contain written instructions for producing your weekly sales recap.

More advanced Skills can also include:

• reference material
• examples
• supporting files
• templates
• executable scripts

Claude can load relevant Skills when a task matches what they are designed to do.

For example, a sales team might create a Skill that tells Claude how to prepare its weekly pipeline report:

• use three specific sections
• pull the relevant CRM information
• highlight deals above a defined value
• flag opportunities that moved backwards
• use the same reporting format each week

The value of the Skill is consistency.

You teach the process once rather than rebuilding the instructions every time.

Creating a Skill

Simple Skills do not require programming knowledge.

At their most basic, they can consist of clear instructions written in Markdown.

More advanced Skills can include scripts and other resources for tasks where certain steps should run consistently.

Claude also offers ways to help you create Skills rather than requiring you to construct everything manually.

A useful starting point is to describe the workflow you want to standardise and provide examples of successful past outputs.

The more concrete the example, the easier it is to define what the Skill should reproduce.

For example:

I want to create a Skill for our weekly sales recap. Here are three reports we’ve used previously. Help me identify the recurring structure and turn it into a reusable process.

When testing a Skill, try it on several realistic tasks rather than relying on a single successful example.

Plugins: Packaging a Broader Workflow

A Skill usually focuses on one reusable process.

A Plugin can package several capabilities together for a broader role or workflow.

Anthropic describes Plugins as bundles that can include:

• Skills
• Connectors
• sub-agents
• other supporting components

For example, a finance Plugin might include separate Skills for:

• variance analysis
• financial modelling
• preparing an investment memo
• recurring reporting

alongside Connectors to the systems those processes depend on.

Plugins can therefore give a person or team a more complete working setup without requiring every component to be configured independently.

Anthropic provides Plugins for different types of work, and installed Plugins can also be customised to better match your organisation’s processes, templates, and examples.

The distinction is useful:

A Skill teaches Claude one repeatable process.

A Plugin packages several capabilities around a broader job or workflow.

Claude in Chrome and Browser Use

Not every business system has a dedicated Connector.

You may use:

• an internal dashboard
• a vendor portal
• a specialist web platform
• an older company system
• a tool available only through a browser

When a Connector is not available, Cowork can use browser capabilities such as Claude in Chrome to interact with supported websites.

For example:

Open our customer-health dashboard, identify the accounts marked red or yellow, then use the relevant customer information from Drive and prepare a one-page briefing for each account.

Claude can use the browser as one part of a larger Cowork task.

This is powerful, but browser access deserves additional caution.

Web pages can contain malicious or misleading instructions, including prompt-injection attempts designed to influence an AI agent.

Be especially careful with:

• financial services
• highly sensitive accounts
• private personal information
• unfamiliar websites
• pages where an incorrect action could have significant consequences

Use Connectors where suitable because they provide a more structured route to information. Use browser interaction when it solves a genuine gap.

Computer Use: Working With Apps That Don’t Have a Connector

Cowork can also interact with applications on your computer through computer use.

Anthropic’s current hierarchy is useful to understand.

Where possible, Claude prefers:

1. Connectors
2. browser interaction
3. direct computer use

A Connector is generally the most structured and reliable option.

Browser interaction fills gaps for web-based tools.

Direct computer use allows Claude to interact with applications by navigating the screen, clicking, and typing.

That makes it possible to work with software that has no dedicated Claude integration.

It also carries greater risk because Claude is interacting with the interface much as a person would.

For sensitive or consequential applications, keep a higher level of supervision.

Claude for Microsoft 365

Claude can also work directly inside Microsoft 365 through add-ins for:

• Word
• Excel
• PowerPoint
• Outlook

This serves a slightly different purpose from Cowork.

Cowork is particularly useful when a task starts with several sources and ends in a new deliverable.

Claude inside Microsoft 365 is useful when the work already lives inside a particular document, spreadsheet, presentation, or email workflow.

Excel

Claude can help:

• analyse data
• create or explain formulas
• investigate errors
• work through scenarios
• make changes within a workbook

PowerPoint

Claude can help create and revise presentations while working with native, editable slide content.

Word

Claude can draft, revise, restructure, and work within documents.

Outlook

Claude can help work with email context and prepare replies or other communication.

Claude can also coordinate context across supported Microsoft 365 apps.

For example, you might begin with a brief in Outlook, use that context while preparing a Word document, build supporting analysis in Excel, and then turn the final recommendations into a PowerPoint presentation.

The distinction is:

Use Cowork when the work spans several sources and you want to delegate the wider task.

Use Claude inside Microsoft 365 when you are working directly within those files and applications.

Working Safely With Cowork

Cowork becomes more useful as you give Claude more access.

That also means setup deserves more care.

Scope File Access Carefully

If a task only needs one project folder, provide that folder.

Do not connect an entire drive simply because it is easier.

The goal is to provide enough access to complete the work without unnecessarily expanding what Claude can reach.

Be Careful With Sensitive Information

Before connecting a folder, service, or website, consider what information it contains.

Avoid exposing sensitive information unless the workflow genuinely requires it and its use fits your organisation’s policies.

Test New Workflows Before Automating Them

Run a new workflow manually before converting it into a scheduled task.

Check:

• what Claude accesses
• how it interprets the instructions
• what it produces
• where it saves or sends anything
• which actions require approval

Once you know the workflow behaves as expected, automation becomes much safer.

Make Consequential Instructions Explicit

Avoid ambiguous language when an action could be interpreted in several ways.

Instead of:

Clean up these files.

say:

Rename these files according to the naming convention below. Do not delete or overwrite any files.

Instead of:

Send the client an update.

say:

Draft the client update for my review. Do not send it.

The more consequential the action, the more useful precise boundaries become.

Monitor High-Stakes Tasks

You don’t need to watch every routine file-processing job from beginning to end.

But if the task touches sensitive information, external communication, purchases, legal documents, or other consequential actions, stay involved.

Human oversight should increase with the stakes.

Be Cautious With Automatic Approvals

Greater autonomy can make repetitive workflows faster.

It can also allow a mistaken interpretation to travel further before you catch it.

Start new workflows with more oversight and relax approvals only when you understand how the process behaves.

Understand Prompt-Injection Risk

Claude may encounter instructions inside websites, documents, emails, or other content that were written by someone else.

Some content can attempt to manipulate an AI system into ignoring your instructions or taking unintended actions.

This is particularly important when Claude can both read external content and perform actions.

Use trusted sources, limit unnecessary access, and be cautious when a task moves from reading information to changing something outside Claude.

Compliance and Organisational Monitoring

Cowork deserves special consideration in organisations with formal compliance or monitoring requirements.

Anthropic provides monitoring options for Team and Enterprise organisations through OpenTelemetry, which can give security teams visibility into areas such as tool calls, file access, approval decisions, and other Cowork activity.

However, monitoring and compliance coverage can differ between local and remote Cowork sessions and does not necessarily map directly to standard Claude audit logging.

If your organisation has formal requirements around:

• audit trails
• regulated workflows
• data retention
• compliance exports
• security monitoring

check Anthropic’s current enterprise documentation before using Cowork for that process.

Do not assume that a workflow is compliant simply because the underlying Claude account is approved for general company use.

Validating a Skill Before You Rely on It

If you create a Skill that will be used repeatedly, test it before treating the process as finished.

Try several realistic prompts that should trigger it.

Then ask:

• Did Claude use the Skill when expected?
• Did the output follow the required structure?
• Were important sections missing?
• Did the tone match?
• Did it apply calculations or rules correctly?
• Does the Skill work with different inputs, not just the example used to build it?

If something is wrong, refine the Skill and test again.

This does not need to become a formal software-testing process.

The aim is simply to find obvious weaknesses before a recurring workflow depends on them.

A Skill that produces one good result once has not necessarily proved that it is reliable.

Common Cowork Mistakes

Treating Cowork Exactly Like Chat

Cowork is most useful when you give Claude an outcome rather than breaking every small step into a separate request yourself.

If you find yourself managing every stage manually, the task may still be functioning like a Chat workflow.

Using Cowork for Something That Needed One Answer

Not every task needs an agentic workflow.

If you simply want five headline options, an explanation of a concept, or feedback on a paragraph, regular Chat will usually be simpler.

Use Cowork when the task benefits from delegation.

Connecting Too Much

More access does not automatically produce better work.

If Claude only needs three files, giving it access to three hundred introduces unnecessary information and risk.

Being Vague About the Deliverable

“Help me with this project” leaves too many decisions undefined.

Name what you want at the end:

• a report
• a spreadsheet
• an email draft
• a presentation
• a comparison
• an organised folder
• a set of recommendations

Then identify the relevant inputs.

Automating Before Testing

A recurring task that is slightly wrong once becomes a recurring task that is repeatedly wrong.

Run the workflow manually first.

Skipping Review Because the Result Looks Professional

A polished presentation can still contain the wrong number.

A beautifully formatted report can still misinterpret the source.

Presentation quality and factual reliability are different things.

Review both.

Giving Claude Too Much Autonomy Too Early

Autonomy is useful once a workflow is understood.

It is not a substitute for understanding the workflow first.

Start with more oversight and expand Claude’s freedom deliberately.

Practical Takeaway

Claude Cowork is designed for delegating complete pieces of work, particularly when a task involves several steps, files, tools, or applications.

The basic workflow is straightforward:

1. Define the outcome.
2. Give Claude access to the information it needs.
3. Set clear boundaries and instructions.
4. Review Claude’s plan and steer when necessary.
5. Check the finished result.
6. Turn successful recurring processes into Projects, Skills, Plugins, or scheduled tasks where useful.

You don’t need to use every Cowork feature at once.

Start with a task that already takes you too many manual steps.

Give Claude a clearly scoped piece of work, see how it handles it, and refine the process from there.

Once that becomes comfortable, features such as Projects, Skills, Connectors, Plugins, scheduled tasks, browser access, and computer use can reduce more of the repetitive work around it.

The goal is not to automate everything.

It is to recognise which parts of your work are worth delegating — and give Claude enough direction, context, and oversight to handle them well.

FAQ

Do I need a paid plan to use Cowork?
Yes. Cowork is currently available on Claude’s paid plans, including Pro, Max, Team, and Enterprise. Availability of individual features can differ by plan and platform, so check Anthropic’s current plan information if you need a specific capability.

Is Cowork only available on desktop?
No. Cowork is now available across desktop, web, and mobile, although some features are still rolling out and certain capabilities involving local files, browser use, or applications on your computer depend on Claude Desktop.

Can Cowork access my entire computer?
Not automatically. Local file access is limited to folders you explicitly make available. However, additional capabilities such as computer use, browser access, Connectors, and Plugins can expand what Claude can interact with, so review those permissions carefully.

Can scheduled Cowork tasks run while my computer is off?
Yes, cloud-based scheduled tasks can run remotely. If a particular task depends on local files or applications on your computer, those local resources will only be available when the necessary desktop connection is active.

Does Dispatch work when my computer is off?
No. Dispatch is specifically designed to let you start work on your desktop from your phone. Your computer must be awake and Claude Desktop must be running while the task uses those desktop resources.

Is a Skill the same as a Plugin?
No. A Skill is a reusable set of instructions and resources for a particular task or process. A Plugin can bundle multiple Skills with Connectors, sub-agents, and other components around a broader workflow.

Should I start with Global Instructions or a Project?
Start using Cowork before worrying about either. Add Global Instructions when you notice yourself repeating preferences that apply across your work. Create a Project when one client, initiative, or ongoing stream of work needs its own context and organisation.

Do Skills require coding?
Not for simple Skills. A basic Skill can be built from instructions and reference material. More advanced Skills can include executable scripts when the workflow benefits from them.

Is Cowork appropriate for legal, financial, or regulated work?
It can assist with preparation, analysis, and drafting, but high-stakes or regulated work requires appropriate human oversight and may have additional compliance requirements. Organisations should check Anthropic’s current monitoring, audit, data-handling, and compliance documentation before incorporating Cowork into regulated workflows.