Week 13 — AI Build Studio · "Automate a Real Task"

Course: Using Artificial Intelligence (AI 101) · Silver Oak University (fictional sample) · Prof. Quinn
Objective: Objective 6 — build a scheduled task, understand the awake-and-app-open constraint, catch an AI over-promise, and verify against official documentation · SLO A (build an agentic workflow with verified outputs) · SLO B (catch AI errors and use agentic tools safely)
Worth 50 points · AI Build Studios group = 15% of the grade · Studio 13
Format: a hands-on build — configure (or carefully design) a real scheduled task, catch the AI claiming it runs while the computer is off, verify against the official docs, and reflect.

Claude Cowork required for this Studio — or a carefully documented design if you are on a free plan. If you don't yet have a paid Cowork plan: complete every design step in writing and use the official documentation to verify your plan. The design skill is what the week teaches; running the task is the bonus. Free tool for the AI-critique step: any approved assistant (ChatGPT, Claude, Gemini, or Copilot).

---

Part 1 — The Build Goal

By the end of this Studio you'll have produced three things:
1. A configured (or carefully designed) scheduled task — a real, specific prompt with a cadence that you've set up in Cowork, or designed in writing with full detail.
2. A documented verification step — you asked an AI to help you configure or describe the task, caught where it over-promised (or confirmed it was correct), and noted the official doc that settles the question.
3. A short reflection on what you'd trust an automation to do — and where you'd still want to be in the loop.

The week's key lesson, stated plainly: an AI assistant that helps you set up a scheduled Cowork task may incorrectly claim the task will run "even if your computer is off" or "on Anthropic's servers." That claim is false — and catching it is this Studio's required step.

---

Part 2 — Pick a Real Repeating Task to Automate

Choose something you actually do regularly (or would want to do regularly) that fits the pattern of knowledge work: summarizing, compiling, researching, organizing. This is more useful and more interesting than an invented scenario.

Good candidates:
- A daily study briefing — each morning, summarize the notes you've added to your class folder since yesterday and list any open action items.
- A weekly class digest — every Sunday, read this week's notes across all your classes and produce a list of concepts to review before the following week.
- A research tracker — once a week, search for recent news on a topic relevant to your major or a project you're working on.
- A task organizer — every Monday, read your to-do folder and compile a prioritized list for the week.

Write your task here: I want Claude to __ on _ [cadence] so that I don't have to ___.

---

Part 3 — Build the Scheduled Task (or Design It in Detail)

If you have a paid Claude plan:

Step 1 — Identify the inputs. What does Claude need to read? Is it files in a connected folder (set up in your Cowork project)? A connected tool like your calendar or a Slack workspace? Make sure the inputs are actually connected.

Step 2 — Write the prompt. Your prompt should include:
- What to read (the input — folder, connector, or tool)
- What to produce (the output — format, length, structure)
- A scope constraint — "base everything only on what's in the files; do not invent or extrapolate" (this prevents hallucination)
- Any formatting rules (e.g., "produce a bulleted list, no longer than 200 words")

Example for a daily study briefing:

"Each morning, read the Markdown files in my AI 101 notes folder that were modified since yesterday. Produce: (1) a 100-word summary of the concepts I added yesterday; (2) a list of any open action items I noted. Base everything only on what's in the files — do not invent content or add details not in my notes."

Step 3 — Create the scheduled task. Use one of the two official methods:
- Method 1: In any Cowork task, type /schedule → follow Claude's prompts → click "Schedule" to confirm.
- Method 2: Click "Scheduled" in the left sidebar → click "+ New task" → fill in name, description, prompt, cadence, and optional folder → click "Save."
- Source: Schedule recurring tasks in Claude Cowork

Step 4 — Note the awake-and-app-open constraint explicitly. In your submission, state: "For this task to run at [time], my computer must be [condition 1] and the [condition 2] must be open." (Fill in the correct conditions — this is the quiz-level fact.)

If you are on a free Claude plan:

Complete every step in writing — your task description, the full prompt you'd use, the cadence, and the specific folder or connector you'd connect. State explicitly: "This task would require a paid plan (Pro, Max, Team, or Enterprise) to run. If I had access, I would [describe the steps you'd take]." The design is the deliverable; the written plan earns full marks.

---

Part 4 — The Required AI-Critique Step (catch the over-promise)

This is the Studio's non-negotiable verification step. It exists because AI assistants — including Claude — sometimes over-promise what scheduled tasks can do.

Run this scenario:
1. Open any approved assistant (ChatGPT, Claude, Gemini, or Copilot — your choice).
2. Ask it to help you set up a scheduled Cowork task. Use a prompt like: "I want to set up a Claude Cowork scheduled task that runs every morning at 7 a.m. to summarize my notes. Will it run reliably even if my laptop is closed or asleep?"
3. Observe the response carefully.

What you're looking for:
- Does the AI correctly state that the task runs only while the computer is awake and the Claude desktop app is open?
- Or does it claim the task will run regardless of computer state, on Anthropic's servers, or even with the laptop closed?

Document what happened:
- Quote or accurately paraphrase what the AI said about when the task would run.
- State whether it was correct or an over-promise.
- If it was an over-promise: correct it — tell the AI it's wrong, cite the official doc (support.claude.com/en/articles/13854387), and ask it to revise. Quote the corrected response.
- If it was correct: state how you verified it (against the official doc) and note that the AI got it right.

Write 3–4 sentences reporting: what the AI said, whether it was accurate, and the verified correct behavior.

The habit all term: the tool drafts or advises; you check against the authoritative source. This week, the authoritative source is Anthropic's own documentation — and the claim about when scheduled tasks run is one the AI might get wrong. Catching it is the skill.

---

Part 5 — Reflection (2–3 sentences)

What kind of task would you actually trust a scheduled automation to handle — and where would you still want to be in the loop before anything is sent, published, or acted on? What did you learn from the AI-critique step that you'll carry into any future automation you build?

---

Part 6 — What to Submit

Submit a single document (or text entry) with:
1. Your task description (what you want automated and why it's a good candidate).
2. Your full prompt (the text you'd use or did use for the scheduled task).
3. The cadence and (if applicable) confirmation that you created the task in Cowork.
4. The awake-and-app-open constraint stated explicitly in your own words.
5. The Part 4 AI-critique write-up — what the AI said, whether it over-promised, and (if so) how you caught and corrected it with the official doc.
6. The Part 5 reflection (2–3 sentences).

Due Sunday, Nov 29, 11:59 p.m. (50 points). Thanksgiving week — plan ahead.

---

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

Students use their own tasks, so deliverables vary. Grade the process and judgment (accurate constraint knowledge, a defensible prompt, a genuine catch or verification), not a specific output.

Model deliverable (illustrative):
- Task: Daily study briefing — each morning, summarize the AI 101 notes folder for concepts covered since yesterday and list open action items.
- Prompt: "Each morning, read the Markdown files in my AI 101 notes folder that were modified since yesterday. Produce: (1) a 100-word summary of the main concepts I added; (2) a bulleted list of any open action items I noted. Base everything only on what's in the files — do not invent or extrapolate."
- Cadence: Daily at 7 a.m.
- Constraint stated: "For this task to run at 7 a.m., my computer must be awake AND the Claude desktop app must be open. If my computer is asleep at 7 a.m., Cowork will skip the task and run it when the computer wakes up."
- Part 4 (model AI-critique): "I asked Claude: 'Will this scheduled task run even if my laptop is closed?' Claude responded: 'Yes, scheduled tasks in Cowork run autonomously on Anthropic's servers — your laptop doesn't need to be on.' This is an over-promise. The official Anthropic documentation at support.claude.com/en/articles/13854387 clearly states: 'Scheduled tasks only run while your computer is awake and the Claude Desktop app is open.' I corrected Claude, cited the doc, and it revised its answer to accurately describe the awake-and-app-open constraint."
- Reflection: Notes that they'd trust automation for read-and-summarize tasks but not for anything involving sending emails or taking action on other people's behalf without review; notes that catching the over-promise was easy once they knew the official constraint.

Why the AI-critique step can't be faked: a student who says "the AI was correct" without quoting the response or verifying against the doc earns the low end of the AI-critique row. The rubric rewards genuine verification — quoting the AI, citing the doc, and either confirming accuracy or catching and correcting an error.

Grading rubric — 50 points

Criterion	Full	Partial	None
Real task + defensible prompt — a genuine repeating task; prompt includes input, output format, and a scope constraint (12)	12	6–9	0–4
Scheduled task built or fully designed — configured in Cowork (or complete written plan with paid-plan note); cadence specified (10)	10	5–8	0–3
Awake-and-app-open constraint — stated accurately and explicitly in the student's own words (8)	8	4–6	0–2
AI-critique / verification step — the AI's claim is quoted or accurately paraphrased; correctly assessed (over-promise or correct); if over-promise, caught and corrected with the official doc citation (15)	15	8–12	0–6
Reflection — a thoughtful, specific take on the trust boundary and what was learned (5)	5	3	0–2

Product-accuracy gate: PASS.

Every Cowork feature in this Studio is verified against official Anthropic documentation:
- Scheduled tasks (awake/app-open constraint, skip behavior, /schedule, the Scheduled sidebar, plan requirements): Schedule recurring tasks in Claude Cowork
- Dispatch (persistent cross-device thread, plan requirements, desktop constraint): Assign tasks from anywhere in Claude Cowork

No features, menu paths, plan details, or behaviors are invented. The "AI over-promise" framed in Part 4 is a real failure mode (an AI incorrectly claiming scheduled tasks run on Anthropic's servers or without the computer being on), explicitly presented as an error to catch — not as correct information.