Week 11 — Practice Exercises · Agents & Claude Cowork I

Course: Using Artificial Intelligence (AI 101) · Silver Oak University (fictional sample) · Prof. Quinn
Week 11 of 16 · Fall 2026 · Objective 5
Format: 5 floor-difficulty exercises · ungraded · recommended by Sunday, Nov 15

Purpose. These exercises lock in the week's concepts before Quiz 11. Work through each one on your own first — then use an approved assistant to check your reasoning. If you get one wrong, figure out why before moving on; the "if incorrect" notes point you toward the right thinking without giving away the answer.

Exercise 1 — Agent or chatbot?

Scenario. You need to rename 200 photo files in a folder so they all follow the format YYYY-MM-DD_description.jpg. You have two options:

• Option A: Open ChatGPT, describe the naming convention, and ask for a script or instructions. You then copy the script, open your terminal, and run it yourself.
• Option B: Open Claude Cowork, connect your photos folder, and give Claude the task: "Rename all .jpg files in this folder to the format YYYY-MM-DD_description.jpg, using the file's creation date and a short description of the image."

Questions:
1. Which option uses a chatbot and which uses an agent?
2. What is the key operational difference between the two approaches?
3. In Option B, what is the ONE most important thing you should do before you let Cowork rename 200 files?

If incorrect on Q1: think about whether the tool is responding (giving you a reply to act on) or acting (taking the steps itself on your files).
If incorrect on Q3: think about what could go wrong if the renaming pattern is wrong — and how you'd catch that before it's applied to 200 files.

Exercise 2 — Cowork vocabulary match

Instructions. Write a one-sentence definition of each term without looking at your notes. Then check yourself.

Key ideas to check against:
- Agent: takes multi-step actions on your behalf (vs. a chatbot that replies once per prompt).
- Project: a persistent workspace with its own instructions, connected folder, context, and memory.
- Connected folder: a local folder on your computer that Claude can read from and write to.
- Task: a prompt you give Claude inside a project that it plans and executes as multi-step work.
- Project memory: Claude retains context from prior tasks in the project and applies it to future ones in the same project.
- Global instructions: standing preferences (tone, format, role context) that apply across every Cowork session.

If you have a term wrong: go back to the "What's inside every project" section of the lecture outline or read the official projects doc at https://support.claude.com/en/articles/14116274-organize-your-tasks-with-projects-in-claude-cowork.

Exercise 3 — Spot the error in the student's description

A classmate wrote this about Cowork. Find and fix the errors:

"Claude Cowork is a browser-based tool available to anyone with a free Claude account. You connect a folder and Claude can read files from it, but it can't write new files — it just gives you the text back in the chat. It has no memory between tasks, so you have to re-explain everything every time. And if you close your browser tab during a task, the task keeps running in the background."

Errors to find (there are four): Write each error and the correction.

If you can't find all four: re-read the "Requirements" section of the Get Started doc (https://support.claude.com/en/articles/13345190-get-started-with-claude-cowork) and the Projects doc. The errors all contradict something stated explicitly in the official docs.

Exercise 4 — Safe-use decision

Scenario. You're a part-time assistant at a small clinic. Your supervisor asks you to use Cowork to batch-summarize 40 intake forms (scanned PDFs) containing individuals' private health information, stored in a shared folder, saving a summary of each to a separate output folder.

Questions:
1. What is the most significant privacy concern with this scenario, and what should you do before running any task?
2. Which permission mode should you use for the first run, and why?
3. After the task completes, what is the first thing you should do before forwarding the summaries to your supervisor?

If incorrect on Q1: think about what kind of private health information is in medical intake forms and whether free-tier AI tools are appropriate for handling it. The Week 15 privacy module covers this in depth — but for now, think about the "billboard test" (would your supervisor be comfortable seeing exactly what you're doing on the front page of a newspaper?).
If incorrect on Q2: the correct answer requires you to prefer caution over speed on a first run with sensitive files.
If incorrect on Q3: the answer is about your verification habit — the same one you practiced in Week 10.

Note: this exercise is for conceptual practice. Do not upload private health records or sensitive personal information to any AI tool without verifying your organization's compliance policies and the tool's data-handling terms.

Exercise 5 — Scenario: write a better project instruction

Setup. You create a Cowork project called "Research Notes" and connect a folder that contains multiple .md files — your notes from reading academic articles. You run this task: "Summarize my notes." Claude produces a single summary that blends all the notes together and doesn't label which article each point came from.

What went wrong? Write one sentence explaining what the task prompt lacked.

Then write a better project instruction (1–3 sentences, written as if you're adding it to the project's Instructions field) that would prevent this problem. A good instruction specifies: the output format, whether summaries should be per-file or combined, and where the output should be saved.

Model (don't read until you've tried): "Produce one summary per .md file in the connected folder. For each file: use the filename as the section header, then write a 5-bullet summary of that file's key points. Save all summaries combined into a single file called research-summaries.md in the same folder. Do not merge content across files."

If you're stuck on what went wrong: the task prompt gave Claude no format instructions, no per-file/combined preference, and no output location. Vague prompts give vague outputs — especially for agentic tasks where there's no back-and-forth to clarify.