Week 11 — Lecture Outline · Agents & Claude Cowork I
Course: Using Artificial Intelligence (AI 101) · Silver Oak University (fictional sample) · Prof. Quinn
Objective covered: Objective 5 — Use Claude Cowork to build agentic workflows — projects, connected folders and file read/write, tasks, and memory.
SLOs touched: A (produce quality results with AI through strong prompting and agentic workflows) · B (evaluate and use AI critically and safely)
Meeting pattern: 2 sessions × 75 min = 150 min. Segment minutes total ~150; scale to your own pattern.
Product-accuracy gate: all Cowork feature claims verified against official documentation (support.claude.com) as of 2026-06-29.
Week at a Glance
| The week's big question | "What changes — in power and responsibility — when an AI can reach into your actual files and act on them?" |
| By the end of the week, students can… | (1) define an AI agent and distinguish it from a chatbot; (2) explain Claude Cowork — the desktop platform, requirements, and how it runs tasks; (3) describe a project (instructions, connected folder, context, memory); (4) explain file read/write (Markdown in, Markdown out); (5) apply the agent safety checklist. |
| Key vocabulary | agent · agentic · chatbot (contrast) · Claude Cowork · task · project · connected folder · project instructions · memory (project-scoped) · global instructions · permission mode (ask vs. act) · least privilege |
| Materials | Slides (Deck 11) · readings + official Cowork docs (H file) · Claude Desktop app (demo machine) · one approved assistant for Tutorial/Discussion/Assignment demo |
| Timing note | 8 segments, ~150 min total. Session 1 = Segments 1–4 (~75 min). Session 2 = Segments 5–8 (~75 min). |
Segment 1 — Hook & the Pivot (8 min) · Session 1 opens
Hook. Write two columns on the board: CHATBOT / AGENT. Put a real scenario in each:
- Chatbot: "Help me summarize my meeting notes." → You paste the notes in, read the reply, copy the summary, paste it somewhere yourself.
- Agent: "Summarize every file in my /meeting-notes folder and save a summary.md to /summaries." → Claude finds the files, reads each one, writes the summary file directly to your folder. You come back to finished work.
Ask: "Which would you rather have for a task you do every week?" Take a show of hands. Then: "Both use the same underlying Claude model. The difference isn't the AI — it's what you gave it permission to do."
Why it matters line: "A chatbot replies. An agent acts. That difference changes everything about how you use it — and how carefully you watch it."
The semester pivot: "For ten weeks we've been prompting. These last six weeks we're building workflows. By the end of November you'll have a real automation running. This week is the foundation."
Segment 2 — What an Agent Is (20 min)
Plain language first. An AI agent is software that takes multi-step actions on your behalf to accomplish a goal. Key word: agentic — the tool has agency, meaning it doesn't just reply, it acts. Compare:
| Feature | Chatbot | Agent |
|---|---|---|
| Execution | One reply per prompt | Plans and executes a sequence of steps |
| File access | You paste content in | Reads from and writes to your actual files |
| Scope | Single conversation turn | Multi-step task; may coordinate sub-tasks in parallel |
| Memory | Within a conversation | Project-scoped: remembers across tasks in a project |
| Supervision | You review reply | You review the plan + the output file |
The classic misconception to defeat: "An agent is just a chatbot with a fancier name." This is the #1 wrong answer on this week's quiz. Stress the difference clearly: a chatbot gives you a draft that you act on; an agent takes the action itself — which means it can be wrong in a much more consequential way.
Why this matters for trust: because an agent writes real files, sends real requests (via connectors), and can make changes that take effort to undo, your verification habit from Week 10 becomes more important, not less. The agent drafts and acts; you still judge and verify.
Quick interaction (2 min): "Give me a real task from your life where an agent saving you the execution steps would be genuinely useful. What would the agent need access to?" (Take 2–3 examples from the room.)
Segment 3 — What Is Claude Cowork? (18 min)
Plain language first. Claude Cowork is the desktop surface of the Claude app where Claude gets file access, runs multi-step tasks in the background, and works inside connected folders. It uses the same agentic architecture as Claude Code — accessible from the Claude Desktop app, without opening a terminal. Source: Get started with Claude Cowork.
Requirements (from the official docs):
- Claude Desktop app (macOS or Windows) — download at claude.com/download. Cowork is not available in the browser or on mobile alone.
- Paid plan: Pro, Max, Team, or Enterprise.
- Active internet connection.
- The Claude Desktop app must stay open while Claude is working. If you close it or the computer sleeps, active tasks stop.
How to access: Open Claude Desktop → find the mode selector → click the Cowork tab to switch to Tasks mode. Describe what you want. Claude shows you its plan; you let it run.
What Cowork can do (from the docs):
- Direct local file access — read from and write to your files, no manual uploads
- Sub-agent coordination — breaks complex work into parallel subtasks
- Professional output formats: Markdown, .docx, .pptx, .xlsx
- Scheduled tasks (covered in Week 13)
- Projects with memory (today's main topic)
- Edit Markdown drafts in place ("highlight → Edit with Claude")
Live demo (~5 min): Open Cowork on the demo machine. Show the mode selector switching to Tasks. Show the project panel on the left. Don't run a long task in class — show the interface and describe what a task start looks like.
Segment 4 — Named Misconception + Quick Review (10 min) · Session 1 closes
Misconception 1: "An agent is just a chatbot." Already covered in Segment 2. Reinforce with a true/false rapid-fire: "Claude Cowork runs in the browser — True or False?" (False — Desktop only.) "A project is the same as a single chat — True or False?" (False — persistent, with memory.) "Cowork can touch your real files — True or False?" (True — that's the whole point.)
Misconception 2: "Cowork can run tasks even when my computer is off." False. From the official docs: scheduled tasks run only while the computer is awake and the Claude Desktop app is open. This is a critical constraint to internalize — relevant for Studio 11 and especially for Weeks 12–13.
Quick interaction: "In one sentence: what's the most important difference between a chatbot and an agent?" Take 3–4 responses. Correct gently where needed.
Session 1 close: "Halfway through. This afternoon/Thursday we get into the structure that makes Cowork useful: projects, connected folders, and the file read/write workflow."
Segment 5 — Projects, Connected Folders & Memory (22 min) · Session 2 opens
Projects in Cowork. A project is a persistent, self-contained workspace with its own files, instructions, and memory. From the official projects doc (Organize your tasks with projects in Claude Cowork):
You create a project by finding Projects in the left navigation panel and clicking +. Three options:
1. Start from scratch — set up a new folder with instructions and files.
2. Import from a Claude project — pull in files and instructions from an existing chat project.
3. Use an existing folder — connect a folder already on your computer, name the project, and add instructions.
What's inside every project:
- Instructions — add tone, format rules, or standing preferences (e.g., "always save outputs as Markdown in the /output subfolder").
- Connected folder — a local folder Claude can read from and write to.
- Context — attach files, link a chat project, or paste a URL for Claude to reference.
- Memory — Claude remembers context from prior tasks in the project and applies it to future tasks. Memory is scoped to the project — what Claude learns in Project A doesn't carry into Project B.
Worked example (narrate live): "I set up a project called 'Weekly Meeting Notes.' I connect my /meeting-notes folder. I add an instruction: 'When I ask for a summary, read all .md files in the input/ subfolder and save a combined summary as summary-YYYY-MM-DD.md in the output/ subfolder.' First task: 'Summarize this week's notes.' Claude reads the files, generates the summary, writes it to the output folder. Next week: same instruction, same project — Claude still knows the format I want because of memory."
Live demo: Navigate to the Projects panel in Cowork. Show the three creation options. Click "Start from scratch" and show the setup form — don't complete it in class, just walk through the fields.
Segment 6 — File Read/Write & the Task Workflow (18 min)
Reading and writing files. The most practical capability in Cowork: Claude reads input files from the connected folder and writes output files to it — no copy-paste, no upload. Markdown is the native format; via built-in skills, Cowork can also output .docx, .pptx, and .xlsx.
The task workflow step by step (from the docs):
1. You give Claude a prompt inside a project.
2. Claude analyzes the request and creates a plan.
3. You see the plan (in "Ask before acting" mode, Claude pauses for your approval at each step).
4. Claude breaks complex work into subtasks, runs code/commands in an isolated virtual machine (VM) on your computer.
5. Outputs are delivered directly to your file system in the connected folder.
6. Progress indicators show what Claude is doing throughout.
Permission modes (doc-verified):
- Ask before acting — Claude pauses for your approval at each step. Recommended when working with new tools, unfamiliar files, or sensitive content.
- Act without asking — Claude works without pausing. Faster, but riskier. Use only when actively supervising and working with trusted files.
- In both modes, Claude requires explicit permission before permanently deleting any files.
Concrete example (write on board or slide):
Connected folder: /Users/yourname/Documents/meeting-notes/
Input file: input/nov-10-notes.md
Task prompt: "Read nov-10-notes.md and write a 3-bullet action-item summary
to output/nov-10-summary.md"
Output file: output/nov-10-summary.md ← Claude writes this
The verify step: after Claude finishes, open nov-10-summary.md. Read it. Did it cover all three agenda items? Did it add any action items that weren't in the notes? Did it write to the right path? This verification moment is the Studio 11 core skill.
Segment 7 — Technology Workflow: Verify the Agent (10 min)
The verify-the-AI moment for agentic tools. Verification is more important here, not less — because the agent acts, and a bad action produces a bad file that you might then act on yourself.
Concrete failure modes to teach:
- Wrote to the wrong subfolder (misread the path).
- Summarized only the first file instead of all files.
- Invented an action item that wasn't in the notes.
- Got the date wrong in the filename.
- Over-confidently completed a task it only partly finished.
The audit checklist for an agent output:
1. Is the file in the right location?
2. Is the filename correct?
3. Does the content match what was in the input files — nothing added, nothing silently dropped?
4. Are there any facts, quotes, or figures that weren't in the original? (Fabrication risk.)
5. Did the agent do the whole task, or stop partway?
Teach the refine-by-instruction habit: if the agent wrote to the wrong location, don't just re-run the task — add a project instruction that specifies the exact output path. Instructions persist; they make the next run better without you supervising every step.
Callback to Week 10: "The verification workflow you built last week — ask for sources, cross-check, ask the AI to critique itself — applies here. The medium changed (it's a file now, not a chat reply), but the discipline is the same."
Segment 8 — Discussion Tease + Hand-off (10 min) · Session 2 closes
Discussion tease. "This week's discussion asks: what changes — for good and for ill — when an AI can read and write your actual files? I want you to take both sides seriously. There are real benefits. There are also real risks — privacy, over-reach, mistakes that propagate. Think it through before you start the dialogue."
Week 11 hand-off:
- Studio 11: "Your First Cowork Project" — create a project, connect a folder, run a task that reads and writes a file, catch the error, reflect. Worth 50 points.
- Assignment 11: agent vs. chatbot distinctions, Cowork concepts explained, a file workflow plan. 100 points, AI coach.
- Quiz 11: 10 auto-graded items, no AI.
Next week: "Week 12 goes deeper — Skills, Connectors (MCP), and live Artifacts. The vocabulary you build this week — project, connected folder, task, memory — is the foundation for all of it."
Final line: "An agent is a tool that acts. The power comes with a responsibility you've already been practicing: verify before you trust."
Instructor FAQ
| Question | Answer |
|---|---|
| "Do I need a paid plan to use Cowork?" | Yes — per the official docs, Cowork requires Pro, Max, Team, or Enterprise. Free plans don't include it. The Studio has a doc-reading + planning alternative for students without access. |
| "Can I use Cowork on my phone?" | Pro and Max users can message Claude from the mobile app while the Desktop app runs the task on their computer. But Cowork itself runs on the Desktop — not mobile-only. |
| "Can Cowork delete my files?" | Claude requires explicit permission before permanently deleting files — you'll see a permission prompt in both modes. But connected folders give Claude read and write access, so it can overwrite files without a deletion prompt. Back up anything precious before running a new task. |
| "What does 'project memory' actually do?" | Claude retains context from prior tasks in the project — e.g., file naming conventions you specified, preferences you stated, errors you corrected — and applies them to future tasks. It doesn't retain memory across different projects. |
| "What if the task stops partway?" | Most common cause: the Claude Desktop app was closed or the computer went to sleep. The app must stay open for tasks to run. Restart the task if needed. |
Scope flag
This week covers Cowork as the students' hands-on setup for the next four weeks. Not covered this week: Skills (W12), Connectors/MCP (W12), Artifacts (W12), scheduled tasks (W13), dispatch (W13), computer use (W14), Claude in Chrome (W14), Claude in Excel (W14). Do not preview these features in detail — the students' first mental model should be clean: project + connected folder + task + memory.
~ Prof. Quinn's edition · Fall 2026 · built with thecoursemaker.com