How to Use Claude Cowork: 7 Capabilities Tutorial

Claude Cowork turned Anthropic's chatbot into a desktop agent back in January 2026. Three months in, the tool handles file management, persistent memory, third-party connectors, reusable skills, project workspaces, browser automation, and scheduled tasks. Same chat interface. Radically different output.

What follows is a walkthrough of each capability with working examples and setup instructions. You will configure Cowork, run a file operation, wire up connectors, build a custom skill, and schedule a recurring task. Budget about half an hour if you follow along.

What you need first: A Claude Pro ($20/month) or Max ($100-$200/month) subscription, the Claude Desktop app on macOS or Windows, and an internet connection.

What You'll Learn

Set up Claude Cowork and run file operations, batch conversions, and PDF splits directly from the chat interface
Configure persistent memory so Cowork learns your preferences, writing voice, and workflow habits over time
Connect Gmail, Notion, Google Drive, and other services through connectors, then build reusable skills from proven workflows
Enable computer use for desktop automation and schedule recurring tasks like daily inbox triage with Dispatch

AI-generated summary, reviewed by an editor. More on our AI guidelines.

Getting Started: Chat vs. Cowork vs. Code

Open the Claude Desktop app. Three tabs sit across the top: Chat, Cowork, and Code.

Chat is the familiar conversational mode. You ask, it answers, you copy the text somewhere useful. File uploads cap at 20 per conversation and 30 MB each. Chat advises. It does not touch your files.

Cowork does the actual work. Grant it a local folder, describe the end state, and walk away. It reads files, builds folders, renames documents, generates spreadsheets, saves everything to disk. Nothing to copy-paste or download after.

Code lives in the terminal. Developers use it for codebases, debugging, shell commands. If you are reading a tutorial about Cowork, you probably don't need the Code tab yet.

The prompt style matters here. Chat prompts spell out tasks: "Review my photos and recommend a folder structure." Cowork prompts pin down the result you want: "Organize these 15 photos into subfolders by topic with descriptive filenames." You paint the finish line. Cowork picks the path.

Initial Setup

Three settings to change before your first session.

Cowork instructions. Under Settings > Cowork, add a guardrail: "Before deleting, overwriting, or renaming any existing file, show me what will change and wait for confirmation." Training wheels. Remove them once you trust the tool.

Memory. Settings > Capabilities. Enable both memory features. Location metadata stays off.

Sandbox folder. Create a folder called Workspace Playground inside Documents. Point Cowork here for your first experiments. Keeps everything contained while you learn.

Click the Cowork tab, select Playground, type a prompt. First time, it asks for folder access. Hit "Always Allow" for folders you plan to revisit.

One gotcha that trips up nearly everyone: Cowork reads files only inside the folder you grant. Drag a file from Downloads into the Cowork window, and it loads into the conversation. But Cowork cannot read that file from disk. You have to physically move it into the granted folder first. Annoying until you internalize it.

Capability 1: File Operations

Cowork creates, edits, organizes, and converts files on your machine. For anyone spending hours wrangling downloads and documents, this alone pays for the subscription.

Expense report from receipt photos

Drop a folder of receipt images into Playground. PDFs, JPEGs, whatever. Then prompt:

I need an expense report from the receipt photos in my Receipts folder.
Give me an Excel spreadsheet with columns for date, vendor, category,
amount, and a totals row. If anything is blurry or unclear, mark it
"verify" in a notes column.

Cowork reads every image with OCR, pulls the data, spits out a formatted .xlsx in your folder. Flagged rows get a "verify" tag. Hundred-plus receipts, a few minutes. Chat cannot even attempt this. The 20-file ceiling kills it before you start.

Batch file conversion

Cowork calls whatever command-line tools you already have installed. LibreOffice handles .docx to PDF. Ghostscript compresses PDFs. ImageMagick converts images. Missing a tool? Cowork asks before installing.

Convert all .docx files to PDF, then move the originals into a
"docx-archive" folder. Compress all PDFs to reduce file size.
Convert all images to PNG with lossless compression. Give me a
before/after summary of total file sizes.

One DataCamp reviewer ran this on a real folder. 21 Word documents became PDFs. 40 PDFs lost 63.7 MB (25.5% smaller). 35 images converted to PNG. Anything over 10 MB got skipped to dodge timeouts. The kind of batch job that used to mean three different websites and an hour of clicking.

Splitting a massive PDF

Some PDFs choke every tool you throw at them. A 400 MB monstrosity that no AI chatbot will ingest in one piece. Tell Cowork:

Split this PDF into separate files, one per chapter or major section.
Use descriptive filenames so I can find what I need at a glance.

It finds natural section breaks, spins up individual files, saves them to your folder. Try doing that through a chat window.

Capability 2: Persistent Memory

Here is where Cowork stops being a file manager and starts acting like it knows you.

Claude Chat stores preferences online with hard capacity limits. Cowork writes memory to files on your computer: CLAUDE.md and Memory.md in the project root. No capacity ceiling. Every correction, every preference you save gets written to disk and loaded into the next session automatically.

Training the memory

Start small. Ask Cowork to summarize a meeting transcript. Read the output. If the structure is wrong, edit the summary yourself. Then:

I've made some changes to your summary. Compare your version with mine
and save those preferences in CLAUDE.md and Memory.md so you remember
them next time.

Cowork diffs the two versions, figures out what you changed, writes those rules to its memory files. Next session, it applies them without being asked.

Email voice extraction

Connect Gmail (next section covers how) and prompt:

Go through my sent emails from the past month. Extract my tone of voice
and save writing style principles you can follow going forward.

Cowork reads your emails, isolates the patterns, saves style rules to Memory.md. After that, every draft it writes matches how you actually sound. Not generic professional-speak. Your actual voice.

And the memory compounds. A week of corrections produces passable results. A month? You get drafts that barely need editing. Not close. Yours.

Capability 3: Connectors

Cowork sees your local folder by default. Connectors push its reach into Gmail, Google Drive, Notion, Slack, and hundreds of other services.

How to connect

Customize sidebar > Connectors > plus icon > Browse. Pick a service, go through the OAuth flow, grant specific permissions. Web connectors talk through browser APIs. Desktop extensions run locally with deeper access.

At minimum, hook up Gmail and Google Calendar to follow the examples in this tutorial.

Cross-referencing meeting notes

Say your team takes notes in Notion but auto-generates transcripts in Google Drive. You can pit them against each other:

Get Implicator.ai in your inbox

Strategic AI news from San Francisco. No hype, no "AI will change everything" throat clearing. Just what moved, who won, and why it matters. Daily at 6am PST.

No spam. Unsubscribe anytime.

Check the Gemini transcript from today's standup against the meeting
notes in Notion. Surface any commitments or action items that didn't
make it into the notes.

Cowork yanks the transcript from Drive, yanks the notes from Notion, runs the comparison, and spits out a list of dropped commitments. Two connectors collaborating while you are somewhere else entirely.

Custom connectors

If a tool is not in the directory, add it through Model Context Protocol. Customize > Connectors > "Add custom connector" and paste the MCP server URL. These run from Anthropic's cloud, not your local machine. Worth knowing if you use internal tools that will never show up in the official catalog.

Capability 4: Skills

Skills work like macros. Define a workflow once, invoke it by name, get identical execution every time.

Building one the right way

Skip the "Create with Claude" button in settings. It produces generic instructions. The approach that works:

Do the work in Cowork first. Ask it to clean up a piece of writing. Review. Give feedback. Refine until the output is right.
Then reverse-engineer it into a skill:

Turn what you just did into a "clear and concise" skill. It should
apply to any text, rewrite for clarity, and provide a changelog showing
what changed.

Answer the clarifying questions Cowork asks about scope and edge cases.
Install. Cowork generates a Skill.md file. Navigate to Customize > Skills > plus icon > Upload a skill.

From that point on, "make this more clear and concise" triggers the skill automatically.

Where skills get powerful

A one-step text cleanup skill is convenient enough. But the same concept scales to ten-step workflows. Weekly leadership reports that pull updates from three teams, standardize the format, and export to PDF. You build the workflow through iteration over a few sessions, then snapshot it into a skill. Drop the raw files in, invoke the skill, get a formatted report back in minutes.

Maintenance reality: After updating a skill, you re-upload the file manually. Skills do not sync across computers, so back them up to Drive. And skills built from real work sessions consistently outperform skills written from cold instructions. Always build the workflow first, codify it second.

Capability 5: Projects

Cowork projects look like Claude Chat projects with one meaningful difference. Cowork writes to its own knowledge files directly.

In Chat, when Claude picks up a new principle about your writing style, it produces an updated knowledge file. You manually add it, manually delete the old version. Tedious. In Cowork:

Codify this principle under the Clarity Partner project: when explaining
a concept, always start with an example first.

Done. Written to the project file. No intermediate steps. Over time, each project accumulates its own rules and preferences that Cowork applies automatically. The projects become opinionated about how you work, which is the whole point.

Keep workflows separate. One project for email, another for reports, a third for research. Each carries its own memory without bleeding into the others.

Capability 6: Browser Extension and Computer Use

Anthropic shipped computer use on March 24, 2026 for macOS. Windows followed ten days later. The feature gives Claude direct control of your desktop through a vision-action loop: take screenshot, analyze what is on screen, decide next action, execute, screenshot again. Repeat until done.

Five-minute setup

Update the Claude Desktop app. Go to Settings > General. Toggle "Computer use" on. macOS asks for Accessibility and Screen Recording permissions. Grant both. That is it. No terminal, no API keys, no configuration files.

The tool hierarchy

Claude does not jump straight to screen control. It picks the fastest available path:

Connectors first. Gmail, Slack, Google Calendar. Direct integrations. Fast and reliable.

Chrome extension second. For web tasks when you have Claude in Chrome installed.

Computer use last. Only when nothing else covers the task. Native desktop apps, proprietary tools, anything requiring GUI clicks.

This hierarchy exists because computer use is slow. The screenshot-analyze-act cycle burns several seconds per interaction. One DataCamp test found that unsubscribing from three email newsletters via the Chrome extension took over 30 minutes. Thirty minutes. For three unsubscribes.

And Anthropic is honest about the limitations. Their documentation says the feature "is still early compared to Claude's ability to code or interact with text." Researchers found a prompt injection attack shortly after Cowork launched that tricked Claude into exfiltrating files. Sensitive apps like trading platforms are blocked by default.

Use computer use for simple, repetitive desktop tasks where no connector exists. Keep it away from anything involving credentials, financial data, or medical records.

Capability 7: Scheduled Tasks

Scheduled tasks run Cowork workflows on a timer. Your machine and the Claude Desktop app need to stay running.

Morning inbox triage

This pulls together every previous capability. Takes a week to calibrate, then pays off daily.

Step 1: Write your email rules. An "Inbox Zero triage workflow" document that describes how you categorize messages, what gets a reply, what gets archived. Save it in a subfolder of your Cowork email project.

Step 2: Wire up Gmail through connectors.

Step 3: Run the triage manually for a week. Cowork reads your inbox, categorizes messages, drafts replies. You review every draft. Correct the ones that miss your tone. Cowork logs those corrections to Memory.md.

Step 4: Create the scheduled task. Open Scheduled Tasks in the Cowork sidebar. Set it for 6 AM daily. Point it at your email project folder.

After that, you wake up to a triaged inbox with drafts that sound like you wrote them. First week is rough. By week three, you rarely edit the drafts.

Dispatch: Phone-to-desktop delegation

Scheduled tasks fire on a clock. Dispatch fires on demand from your phone.

Open the Claude mobile app, type what you need, and Claude handles it on your desktop while you are elsewhere. A push notification lands when the work wraps. Things people actually use it for: pulling weekly metrics into a spreadsheet, exporting a meeting deck to PDF and attaching it to a calendar invite, batch-resizing a folder of screenshots.

One requirement nobody reads until they get burned: your computer has to stay awake. The screen goes dark, the task dies mid-run.

Common Mistakes and How to Avoid Them

Wrong folder access. Cowork reads only what is inside the folder you grant. Dragging a file into the conversation from elsewhere loads it visually but does not give Cowork disk access. Move files into your granted folder before you start. Build the habit early.

Chat-style prompts. "Review my files and suggest a naming convention" gets you a suggestion. "Rename all files using YYYY-MM-DD-description format" gets you renamed files. Cowork prompts define the outcome, not the task. If you keep getting advice instead of action, that is the fix.

Skipping the feedback phase. Cowork gets dramatically better with corrections, but most people never bother. When the output is close but slightly off, edit it yourself and tell Cowork to save the delta. Five early corrections outweigh fifty later ones.

Overusing computer use. Screen control is a last resort, not a shortcut. Connectors and Chrome run 5-10x faster for tasks they can handle. If you find Cowork clicking through a browser for something Gmail or Slack could do natively, you have a connector problem, not a Cowork problem.

Cold-writing skills. The "Create with Claude" shortcut in settings produces mediocre skills. Work through the actual process first, iterate three or four times, then tell Cowork to turn the conversation into a skill. Reverse-engineering from real work beats speculative instruction writing every time.

What Works and What Breaks

Cowork handles file operations well. Organizing, converting, renaming, splitting. Repetitive workflows too: weekly reports, email triage, receipt data extraction. Tasks that pull from multiple sources through connectors also land solidly.

Where it breaks: complex spreadsheets with merged cells or presentation-style layouts. The xlsx parser expects tidy columns and falls apart on anything else. Chrome automation technically works but crawls. And computer use remains a research preview with documented security gaps.

Start with one folder and one task. Build memory through honest feedback. Add connectors as you need them. Create skills from workflows you have already proven work by hand. Cowork rewards patience over ambition. Your setup three months from now will barely resemble what you build this week, and that is how it should work.

Frequently Asked Questions

Do I need a paid subscription to use Claude Cowork?

Yes. Cowork requires a Claude Pro ($20/month) or Max ($100-$200/month) subscription. The free tier does not include Cowork access. You also need the Claude Desktop app installed on macOS or Windows.

What is the difference between Cowork and Claude Code?

Cowork runs through the familiar chat interface and targets non-developers who need file management, workflow automation, and task execution. Code runs in a terminal and targets developers building software. Both share the same underlying agent architecture.

Can Cowork access files outside the folder I grant permission to?

No. Cowork reads only files inside the folder you explicitly grant access to. Dragging a file from elsewhere into the Cowork window loads it into conversation context but does not give Cowork disk-level access to that file.

How does Claude's computer use feature work?

Claude takes a screenshot of your screen, analyzes what is visible, decides the next action (click, type, scroll), executes it, then takes another screenshot. This vision-action loop repeats until the task is complete. It works with any installed app but runs slower than direct connectors.

Is it safe to let Claude control my desktop?

Anthropic blocks sensitive apps by default and requires per-app permission. However, the feature is still a research preview. Researchers demonstrated a prompt injection vulnerability shortly after launch. Avoid using computer use with financial, medical, or credential-related applications.