Week 11 · AI Build Studio

Week 11 — AI Build Studio · "Your First Cowork Project"

Using Artificial Intelligence · AI 101 Fall 2026 · Prof. Quinn Fictional sample

Course: Using Artificial Intelligence (AI 101) · Silver Oak University (fictional sample) · Prof. Quinn
Objective: Objective 5 — install/set up Claude Cowork desktop; create a project; connect a folder; run a task that reads an input file and writes an output file; verify the output and catch errors · SLO A (agentic workflows) · SLO B (verify and use AI safely)
Worth 50 points · AI Build Studios group = 15% of the grade · Studio 11
Format: hands-on build — you'll create a real Cowork project, run a real file-read/file-write task, catch an error in the output, and write a verification report.

This is the course's signature weekly component. Every instructional week has one Studio — a real thing to build, a required step where you verify and improve the AI's work, and a short reflection.

Cowork access required. Claude Cowork requires the Claude Desktop app and a paid plan (Pro, Max, Team, or Enterprise). If you don't have access, use the Alternative Path in Part 6 — it earns full credit and is just as valuable.

Product-accuracy gate: all Cowork features described in this Studio were verified against the official Anthropic documentation before shipping. Verified docs: Get started with Claude Cowork · Organize your tasks with projects in Claude Cowork · Use Claude Cowork safely.

Part 1 — The Build Goal

By the end of this Studio you'll have produced three things:
1. A working Cowork project with a connected folder and project instructions (or a detailed plan for one, if using the alternative path).
2. An output file Claude wrote to your connected folder based on an input file you provided.
3. A short verification and error-catch report naming at least one thing Claude got wrong, wrote to the wrong place, or mishandled — and what you'd do to fix it.

This is the whole week in miniature: use an agent effectively (set up the project right) AND judge it well (verify the output, catch the error).

Part 2 — Setup: Install and Access Claude Cowork

If you already have Claude Desktop with Cowork enabled, skip to Part 3.

Step 1: Download Claude Desktop.
Go to https://claude.com/download and download the app for your system (macOS or Windows). Install it.

Step 2: Check your plan.
Cowork requires a paid plan (Pro, Max, Team, or Enterprise). Open Claude Desktop, go to your account settings, and confirm your plan. If you're on a free plan, go directly to the Alternative Path in Part 6.

Step 3: Run the readiness check (optional).
If you want to confirm your system will support Cowork before installing, the official docs provide readiness-check links for macOS and Windows: see https://support.claude.com/en/articles/13345190-get-started-with-claude-cowork.

Step 4: Access Cowork.
Open Claude Desktop. Look for the mode selector at the top. Click the Cowork tab (or "Tasks" — the exact label may vary by version; see the official Getting Started doc). You're now in Cowork mode.

Part 3 — Create Your First Project

Step 1: Create a dedicated folder on your computer.
On your desktop or in Documents, create a folder called cowork-studio-11. Inside it, create two subfolders: input/ and output/. This is your workspace — Claude will read from input/ and write to output/.

Step 2: Create an input file.
Inside input/, create a plain text or Markdown file called notes.md. Write at least 6–8 sentences of genuine notes — use your actual notes from a class, a meeting, a book you're reading, or anything real you have. Do not use dummy "lorem ipsum" text — this should be real content so the summary is meaningful and verifiable.

Step 3: Create a Cowork project.
In Claude Desktop, find Projects in the left panel and click +. Choose "Start from scratch" (or "Use an existing folder" — either works). Name it something like "Studio 11 – Notes Summarizer". When prompted, connect your cowork-studio-11 folder.

Step 4: Add project instructions.
In the project's Instructions field, type something like:

"Read any .md files in the input/ subfolder. Produce a structured Markdown summary for each file: use the original filename as the header, list the 5 key points as bullets, and identify any action items at the end. Save each summary as [filename]-summary.md in the output/ subfolder. Do not merge content across files. Do not add information that is not in the original file."

Save the instructions. This tells Claude exactly what to do — and the "do not add information" line is the constraint you'll use to catch any fabrication.

Part 4 — Run the Task

Step 1: Start a new task in your project.
Make sure you're inside the project you just created. In the task input, type:

"Summarize all .md files in the input/ subfolder per the project instructions and save each summary to the output/ subfolder."

Step 2: Choose your permission mode.
Since this is your first task and you're learning the workflow, set the mode to "Ask before acting" (the doc recommends this for new tasks and unfamiliar files). This lets you see Claude's plan before it executes each step.

Step 3: Review the plan, then let it run.
Claude will show you what it's about to do. Read the plan: does it say it will read from input/notes.md and write to output/notes-summary.md? If something looks wrong, stop and revise the task or the instructions before proceeding. If the plan looks right, approve and let it run.

Step 4: Watch the progress.
You'll see progress indicators as Claude reads the file, generates the summary, and writes the output. This is the agentic execution in action — multi-step work, not a single reply.

Part 5 — Verify the Output (required — this is the core skill)

Step 1: Open the output file.
Navigate to your output/ subfolder. Open notes-summary.md (or whatever Claude named it). Read the entire file. Do not assume it's correct because Claude produced it confidently.

Step 2: Verify against the original.
Go back to your input/notes.md. Compare the summary against the original. Run this checklist:
- [ ] Location: Is the file in output/, not somewhere else?
- [ ] Filename: Does it follow the naming convention from the instructions?
- [ ] Coverage: Does the summary cover the right content — not just the first paragraph?
- [ ] No fabrication: Did Claude add any bullet points, facts, or action items that weren't in your original notes? (This is the key catch — even a factual summary can fabricate.)
- [ ] Action items: If there were real action items in your notes, did Claude find them? Did it invent any?
- [ ] Complete task: Did Claude summarize the whole file, or did it stop partway?

Step 3: Catch at least one error (required).
For this Studio, you must find and report at least one thing Claude got wrong, overstated, under-covered, or mislocated. If everything looks genuinely perfect (rare — but possible), write exactly how you verified each claim and why you're confident nothing was fabricated.

Common errors in this type of task:
- Added a bullet point that wasn't in the notes (fabrication).
- Missed the last section of the notes.
- Named an action item that was actually just a question in the notes.
- Got the filename slightly wrong.
- Wrote the file to the project root instead of output/.

Step 4: Write a better project instruction (required).
After catching the error, write one new or revised project instruction that would prevent it from happening again. Example: if Claude fabricated an action item, add to the instructions: "Only list action items that are explicitly stated in the notes with words like 'to-do,' 'action,' 'follow up,' or an imperative verb. Do not infer or generate action items."

Part 6 — Alternative Path (no Cowork access)

Use this path if you don't have a paid Claude plan or the Claude Desktop app. It earns full credit.

Instead of running a live task, complete this planning and analysis exercise:

Design the project (as in Parts 2–3): describe in writing the folder structure you would create, the project name, the exact project instructions you would add, and the task prompt you would use. Be specific — include file paths, naming conventions, and the constraint that prevents fabrication.
Read the official docs: read both required docs from H-readings: Get started with Claude Cowork and Organize your tasks with projects in Claude Cowork. List three specific things you learned that you would apply to your project setup.
Error prediction: based on the lecture and the docs, identify three specific things that could go wrong in the file-read/write task (wrong output location, fabricated content, incomplete coverage, etc.) and write a project instruction or verification step for each.
Verify a Cowork claim: using one of the required docs, find one claim about Cowork that you initially assumed was true but was more nuanced or more limited than you thought. Describe the claim you held and what the docs actually say.
Reflection (same as Part 7 below).

Part 7 — Reflection (required — 2–3 sentences)

Answer one of these: What surprised you most about how the agent carried out the task (or, for alternative path, about what you read in the docs)? What would you do differently in your project instructions next time? What does this Studio change about how you think about the difference between prompting a chatbot and directing an agent?

Part 8 — What to Submit

Submit a single document or text entry with:
- Your project setup description (project name, folder structure, the instructions you wrote).
- Your task prompt (the exact text you used or would use).
- The output file (paste the content of notes-summary.md) — OR for the alternative path, your project design document.
- Your Part 5 verification report: the checklist (checked), at least one error caught (described specifically), and the new/revised project instruction you wrote to prevent it.
- Your Part 7 reflection.

Due Sunday, Nov 15, 11:59 p.m. (50 points).

Instructor answer key and model deliverable — REMOVE BEFORE PUBLISHING TO STUDENTS

Students use their own notes, so deliverables vary. Grade the process (project setup quality, task precision, verification rigor, error catch, reflection), not a specific output. The model below shows what full credit looks like.

Model (illustrative):
- Project name: Studio 11 – Notes Summarizer
- Folder: cowork-studio-11/input/ contains notes.md (8 sentences of real class notes); output/ is the target
- Instructions: "Read all .md files in input/. Produce a structured Markdown summary for each: filename as header, 5 key-point bullets, any action items explicitly stated in the file (not inferred). Save as [filename]-summary.md in output/. Do not add information not in the original file."
- Task prompt: "Summarize all .md files in input/ per project instructions and save to output/."
- Output verification: file is in the right location and named correctly; 5 bullets cover the right sections; but bullet 4 says "Action item: schedule a meeting" when the original note said "maybe schedule a meeting?" — an inferred/fabricated action item. Catch: Claude over-committed a tentative note as an action item. Fix added to instructions: "Only list action items explicitly marked with 'action,' 'to-do,' 'follow up,' or a clear imperative. Do not infer action items from tentative language ('maybe,' 'could,' 'might')."
- Reflection: "I didn't expect Claude to turn a question mark into a to-do. It made me realize that 'do not add information' is too vague — I needed to be specific about what counts as an action item."

Why the verification step can't be faked: a student who submits an unchecked output with no error caught earns the low end of the verification row. The rubric rewards the specific catch and the specific fix, not the AI's output quality.

Product-accuracy gate: PASS. Every Cowork feature, requirement, and menu path described in this Studio was verified against the official Anthropic documentation (support.claude.com) as of 2026-06-29. Verified docs: Get started with Claude Cowork · Organize your tasks with projects in Claude Cowork · Use Claude Cowork safely. No feature, plan tier, setting, or constraint is invented without a documentation source. The alternative path ensures students without Cowork access can earn full credit through equivalent doc-verified learning.

Grading rubric — 50 points

Criterion	Full	Partial	None
Project setup — project name, folder structure, and project instructions that specify format, output location, and the no-fabrication constraint (10)	10	5–8	0–4
Task execution (or design) — task prompt is specific enough to run without ambiguity; permission mode chosen correctly; progress noted (10)	10	5–8	0–4
Verification checklist — all checklist items addressed; at least one specific error caught and described precisely (not just "it made a mistake") (15)	15	8–12	0–7
Revised instruction — writes one new or revised project instruction that specifically prevents the error caught (10)	10	5–8	0–4
Reflection — a genuine takeaway about directing an agent vs. prompting a chatbot (5)	5	3	0–2

~ Prof. Quinn's edition · Fall 2026 · built with thecoursemaker.com