Teacher’s Edition
Module 10
Claude Cowork: Your Desktop Operator
Claude moves from browser tab to desktop colleague. Cowork reads your files, creates documents, saves them where you specify. The shift from "I type, it responds" to "I direct, it works."
Charter Oak Strategic Partners · Claude Mastery Program · Version 1.0 · Confidential · Not for distribution to participants
Cowork is a feature of the Claude Desktop app (macOS and Windows) that brings Claude Code’s agentic capabilities to knowledge work. It is available on paid plans (Pro, Max, Team, Enterprise) and is currently a research preview. Where regular Claude chat works one message at a time, Cowork takes multi-step tasks and executes them autonomously: reading files from your computer, writing code, creating documents, saving output to your file system, and running for extended periods without supervision.
The architectural difference matters for teaching: Cowork runs in a virtual machine on the participant’s computer. Code executes in an isolated sandbox, but file changes write to the real file system. This means Claude can create real files, edit real documents, and save real output — but the execution environment is contained. Participants need to understand this security model to trust Cowork enough to delegate real work to it.
This distinction will come up immediately. Prepare for it. Chat is the conversational interface everyone has been using in Tiers I and II: you type, Claude responds, one turn at a time. Projects are persistent workspaces with uploaded knowledge and custom instructions that Claude uses as background context. Cowork is the autonomous execution layer: you describe an outcome, and Claude plans, executes multiple steps, reads and writes files, runs code, and delivers finished work. The shift is from “Claude answers my questions” to “Claude completes my work.”
Concrete example for the room: “In Chat, you would paste a CSV and ask Claude to summarize it. In a Project, you would upload that CSV to the knowledge base and have ongoing conversations about it. In Cowork, you would point Claude at a folder containing the CSV and say: read this data, create a summary report with charts, write speaker notes for a presentation, and save both files to the output folder. Claude does all of it while you get coffee.”
Opening: The Cowork Interface — 20 minutes
demo-data/module-10/cowork-sprint-brief.md— Full brief with setup instructions and 100-point scoring rubric.demo-data/module-10/quarterly-sales-messy.csv— Q4 sales data with intentional quality issues for the sprint exercise.demo-data/module-10/sample-interview-transcript.txt— Interview transcript with Maria Flores about operations improvements.demo-data/module-10/employee-handbook-excerpt.txt— HR policy excerpt for the QBR presentation exercise.“Open Claude Desktop. You will see two options: Chat and Cowork. Select Cowork.”
Show the screen to the room. Walk through every element:
1. The folder selector. Cowork asks you to select a folder. This is the folder Claude can read from and write to. Everything Claude creates goes here unless you specify otherwise. “Think of this as your project folder. You would not give a new employee access to your entire computer. Same principle.”
2. The conversation pane. This is where you describe the task and see Claude’s progress. Unlike Chat, Claude shows you its plan and the steps it is executing. You can see it thinking, reading files, writing code, and creating output.
3. Global instructions. Navigate to Settings and show where global instructions are set. “These are standing instructions that apply to every Cowork session. Your preferred tone, your role, your output format preferences. Set once, applies everywhere.”
4. Folder instructions. Show that Cowork can pick up a CLAUDE.md file from the selected folder. “These are project-specific instructions. Different from global instructions. If you are working on a financial report, the folder instructions might say: always use fiscal year dates, round to thousands, include footnotes for every data source.”
Before the sprint, demonstrate each of these so participants see them in action:
File operations. Claude reads files from the selected folder, creates new files, and saves them. Point to a CSV. Ask Cowork to “read quarterly-sales.csv and tell me the top three products by revenue.” Show the response referencing actual data from the file.
Document creation. Ask Cowork to “create a one-page executive summary of this data as an HTML file.” Show the file appearing in the folder. Open it in a browser.
Multi-step execution. Give Cowork a compound task: “Read the sales data, identify the three biggest quarter-over-quarter changes, write a paragraph explaining each one, and save the analysis as analysis.md.” Show Claude planning the steps, executing them in sequence, and producing the output.
Code execution. Cowork can write and run code inside its sandbox. Ask it to “create a chart showing monthly revenue trends and save it as revenue-chart.html.” The code runs in the sandbox; the output file is saved to the folder.
Extended tasks. Mention that Cowork can run for extended periods. “If you give Cowork a complex task — research, multiple documents, cross-referencing — you can step away and come back. This is not a chat that times out after a single response. It keeps working until the task is done.”
Cowork supports scheduled tasks. Type /schedule in any Cowork session, or click “Scheduled” in the left sidebar. You can create tasks that run automatically on a recurring basis or on demand. The constraint: scheduled tasks only run while the computer is awake and Claude Desktop is open. This is worth stating explicitly, because participants will assume it runs like a server.
Example for the room: “Every Friday at 4 PM, Cowork reads the latest CRM export, calculates the week’s pipeline metrics, and saves a formatted report to the shared drive. You set it up once. It runs every week. The caveat: your laptop needs to be open and Claude Desktop needs to be running at 4 PM on Friday.”
Cowork supports plugins that customize how Claude works for specific roles, teams, or companies. Each plugin bundles skills, connectors, slash commands, and sub-agents into a single package. Anthropic has released plugins for financial services (Bloomberg, FactSet, S&P), and the partner ecosystem includes Notion, Figma, and Atlassian integrations. For Team and Enterprise plans, admins can provision plugins organization-wide.
For this module, plugins are worth mentioning but not the focus. The focus is the core Cowork capability: autonomous file-based work. Plugins extend it; they do not define it.
On macOS, Cowork can interact with Excel and PowerPoint through Office add-ins. Claude can open a spreadsheet, analyze data, create a chart, and move that chart into a PowerPoint presentation — all within the same Cowork session. This is a research preview and is not yet available on Windows.
If your participants use Microsoft Office heavily, demonstrate this. If not, mention it as a capability that is coming and move on.
Setup and Prerequisites — Before the Session
Every participant needs Claude Desktop installed with a paid plan (Pro, Max, Team, or Enterprise). Verify this before the session. Cowork is not available on free plans. If even one participant cannot access Cowork, pair them with someone who can.
Pre-load the sprint folder on each machine with three source files: quarterly-sales-messy.csv, sample-interview-transcript.txt, and employee-handbook-excerpt.txt. All three are included in this module’s demo data. The folder should also contain the scoring rubric (cowork-sprint-brief.md).
Test one machine end-to-end the morning of. Open Claude Desktop, select the sprint folder, give it a task, verify it reads the files and produces output. Technical failures during a timed exercise destroy momentum. If Cowork is running an update (“Setting up Claude’s workspace”), it may take a few minutes on first launch. Account for this.
Permissions check: Cowork inherits the same permission settings as Chat. File access, MCP connectors, and code execution should be enabled in Settings. Walk participants through this if needed.
The Sprint — 60 minutes
“You have sixty minutes. Open Claude Desktop. Select the sprint folder. Inside you will find three files: a sales CSV, an interview transcript, and an employee handbook excerpt.”
“Your deliverable: a quarterly business review presentation. Six slides minimum. A one-page executive summary. Speaker notes for every slide. Use all three source files meaningfully. Do not just mention them. Extract real numbers from the CSV. Pull real quotes from the transcript. Reference real policies from the handbook.”
“Start with a comprehensive brief. Tell Cowork everything it needs to know about the task in your first message: the deliverable, the format, the sources, the audience, and the quality bar. The clearer your opening brief, the better the output. This is delegation, not chat.”
“The clock starts now.”
Put a visible timer on the projector. The countdown creates urgency. That urgency is the teaching tool — it forces participants to delegate rather than micromanage.
Participants who have spent eight modules in Claude’s chat interface will instinctively try to write every sentence themselves and use Cowork as a formatting tool. This defeats the purpose. When you see it, intervene directly: “Tell Cowork what you want the slide to say. Give it the data. Let it write. You edit the result. You do not draft.”
The sprint is designed to break the habit of doing Claude’s work for it. Participants who try to draft every slide themselves will run out of time. Participants who delegate clearly will finish with time to spare. The time pressure makes this lesson self-teaching.
Cowork cannot access the folder: Click the folder icon in the sidebar and re-select the folder manually.
Claude produces an error reading a file: Clear the conversation and start a new Cowork session. Sometimes restarting the session resolves file access issues.
“Setting up Claude’s workspace” message: This means Cowork is updating. Wait 1-2 minutes. This is expected on first launch.
Usage limit hit: Cowork consumes more usage than standard Chat because multi-step tasks are token-intensive. If a participant hits their limit, pair them with someone who has capacity remaining. For future sessions, recommend Max plans for heavy Cowork use.
Claude stopped working mid-task: The most common cause is the computer going to sleep or Claude Desktop being minimized/closed. Ensure energy settings prevent sleep during the sprint.
Have a printed troubleshooting checklist at each table: folder selected? Files in the right location? Claude Desktop latest version? Paid plan active? Code execution enabled in Settings?
Walk the room every ten minutes. At the 30-minute mark, do a checkpoint: “Show me your progress.” Anyone with fewer than two slides at this point is stuck. Unstick them by looking at their conversation history and identifying where Cowork went off track. The most common fix: “Start a new conversation. Be more specific in your opening brief. Tell Cowork the full scope upfront — all six slides, the executive summary, the speaker notes, which files to use for what.”
At the 45-minute mark, announce: “Fifteen minutes left. If your presentation is mostly done, start polishing. If it is not, focus on getting a complete draft rather than a perfect partial.”
Security and Permissions — 5 minutes (woven into setup or debrief)
Cowork runs code in a virtual machine on the participant’s computer. This provides isolation: code cannot escape the sandbox. However, Claude can make real changes to files in the selected folder. This is the trade-off: Claude needs real file access to be useful, but that access means it can overwrite or delete files.
Practical guidance for participants: “Select a dedicated working folder, not your entire Documents directory. Treat Cowork like a new hire with access to one project folder. Review what it creates before sending it to anyone. Back up important files before letting Cowork modify them.”
For Team and Enterprise plans, admins can control MCP connectors, file access, and other permissions centrally. Mention this for participants who will need to roll Cowork out to their teams.
Review and Scoring — 15 minutes
The sprint brief includes a detailed rubric. Walk through it before scoring begins:
Data integration (25 points): Does the presentation use real numbers from the CSV? Not made-up numbers, not rounded estimates — actual figures traced to the source file. Deduct points for any number that cannot be verified against the original data.
Narrative quality (25 points): Does the story connect data to decisions? A good QBR does not just report numbers. It explains what the numbers mean and recommends what to do about them. Look for a clear recommendation in the executive summary.
Source coverage (20 points): Were all three files used meaningfully? A sales number from the CSV, a direct quote from the interview transcript, a policy reference from the handbook. Mentioning a file is not using it. Extracting specific content from it is.
Presentation quality (20 points): Are the slides clean? Are the speaker notes useful (not just a repeat of the slide text)? Is the executive summary executive-ready (one page, leads with the conclusion, includes a recommendation)?
Completeness (10 points): Six slides, speaker notes for every slide, one-page executive summary. Missing any deliverable loses points.
“Score yourself first using the rubric. Be honest. Then swap with a neighbor and score each other. Where you disagree, discuss why.”
Give five minutes for self-scoring, five minutes for peer review, five minutes for group discussion.
After scoring: “Hands up: who scored above 80? Above 60? Below 40?” Then: “What worked? What required extra turns? What would you do differently in your opening brief?”
The most common finding: participants who gave Cowork a detailed brief at the start produced better results than those who fed it one instruction at a time. The lesson: treat Cowork like a colleague you are delegating to. The clearer the delegation, the better the work.
“What surprised you about working with Cowork versus Chat?”
“Where did you have to override Claude’s output? What does that tell you about where human judgment still matters?”
“How would you use this in your actual job next week?”
Be transparent about what Cowork cannot do yet. This builds credibility and prevents frustration when participants try it on their own:
No memory across sessions. Each Cowork session starts fresh. Claude does not remember what it did in a previous session. Workaround: save important context in the folder as a text file that Claude reads at the start of each session.
No session sharing. You cannot share a Cowork session with a colleague the way you can share a Chat conversation.
Scheduled tasks require the computer to be awake. Unlike a server-based automation, Cowork scheduled tasks only run when Claude Desktop is open and the computer is not asleep.
Usage is higher than Chat. Multi-step autonomous tasks consume more tokens. Participants on Pro plans may hit usage limits faster with Cowork than with Chat.
| Segment | Activity | Time |
|---|---|---|
| Interface Tour | Cowork walkthrough: folder, instructions, capabilities | 20 min |
| Sprint | Build the QBR presentation from three source files | 60 min |
| Review | Self-score, peer review, group debrief | 15 min |