Week 12 · AI Build Studio

Week 12 — AI Build Studio · "Skill, Connector, or Artifact?"

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 — build an agentic Cowork workflow using skills, connectors, or live artifacts; verify the result; catch an over-claim · SLO A (produce a real result with AI) · SLO B (verify AI feature claims against docs)
Worth 50 points · AI Build Studios group = 15% of the grade · Studio 12
Format: a real hands-on build — you'll use (or design) a Claude Cowork feature, verify the result, catch an over-claim, and reflect on what you learned about both the tool and the limits of what AI says about itself.

This is the course's signature weekly component. Every instructional week has one Studio — a real thing to build, a required verification step, and a short reflection. Links below are to official sources only, each verified live before this module was published.

Part 1 — The Build Goal

By the end of this Studio you'll have:
1. Used (or carefully designed) one of three Cowork features — a built-in skill, a low-risk connector, or a live artifact — and documented what you did.
2. Verified the result against official documentation, checking whether the feature did what it was supposed to do and what the docs say it can do.
3. Caught at least one over-claim — a claim about Cowork that an AI made, a blog post asserted, or you found in the wild that doesn't match the official docs — and corrected it with evidence.
4. Written a short reflection on what you learned about feature distinctions and product-accuracy verification.

This Studio is the course's product-accuracy discipline in full: use a real tool, check what the docs say, catch the error, fix it.

Part 2 — Choose Your Build Path

Pick ONE of the three options below. Each path is equally valid; pick the one that fits your current setup and comfort level.

Option A — Use a Built-in Skill (Recommended for everyone)

You need: Claude Cowork desktop (installed in Week 11). Free.
Download if needed: https://claude.ai/download

What to do:
1. Open Claude Cowork and go to an existing project (or create a simple new one with a connected folder).
2. Give Claude a task that uses a built-in document skill — the most accessible are:
- docx (Word): "Take these notes and create a Word document with three sections: Key Points, Action Items, and Next Steps." Paste in a short list of real notes (about anything — a class, a project, a meeting you'd be okay documenting).
- pptx (PowerPoint): "Create a 5-slide PowerPoint outline on [any topic from your major or life], with a title, three content slides, and a conclusion."
- xlsx (Excel): "Create a simple Excel workbook tracking [something you track — study hours, spending, to-dos] with headers and 5 example rows."
3. Run the task. Open the generated file. Confirm it's a real, properly formatted document (not just text in the chat).
4. Document: screenshot or copy the task prompt you used + a description of the generated file.

Verification step: go to https://code.claude.com/docs/en/skills and read the skills documentation. Confirm: is what you did consistent with how skills actually work? Write 1–2 sentences describing what you verified.

Option B — Connect a Low-Risk Connector (If you have Cowork and a willing app)

You need: Claude Cowork desktop + one account to connect (see below for suggested low-risk options).

Low-risk connector suggestions (from the official connectors directory — read the permission page before connecting):
- GitHub (read-only to your own public repos) — pull a list of your open issues into a Cowork task.
- Google Calendar (read-only, specific calendar only) — pull today's or this week's events into a morning briefing task.
- Notion (read-only to a specific page or database you own) — pull notes from a personal Notion page.

What to do:
1. Open the connectors directory in Claude Cowork (find it via the Customize menu in the sidebar, or at: https://support.claude.com/en/articles/11176164-use-connectors-to-extend-claude-s-capabilities).
2. Read the connector's detail page before connecting: what can it read? what can it write? what permissions does it ask for?
3. Connect the connector with only the minimum permissions needed (least privilege — read-only if the task is a read task).
4. Run a task using the connector (e.g., "give me a summary of my open GitHub issues" or "what's on my calendar this week?").
5. Document: what connector you chose, what permissions you granted, and what Claude produced.

Verification step: go to https://support.claude.com/en/articles/11596036-anthropic-connectors-directory-faq and read the FAQ. Confirm: are the permissions you granted consistent with least-privilege principles? Did Claude's output reflect real data from the connector, or did it fill in something from its training? Write 1–2 sentences describing what you verified.

If you're not comfortable connecting a real account: skip to Option C or the Alternative Path below.

Option C — Design and Document a Live Artifact

You need: Claude Cowork desktop + at least one connector set up (from Option B, or from Week 11 setup).

What to do:
1. Ask Claude to build a live artifact — a persistent, refreshing view — from data in one of your connected apps. Examples:
- "Build me a live view of my open GitHub issues that refreshes from my GitHub connector."
- "Create a live morning briefing artifact that pulls this week's calendar events."
2. Open the artifact in its own tab. Confirm it is distinct from the chat thread.
3. Document: the artifact you created, the connector it uses, and what "live" means for this artifact (when and how it refreshes).

Verification step: go to https://support.claude.com/en/articles/14729249-use-live-artifacts-in-claude-cowork and read the official article. Confirm: does your artifact behave as described? Does it actually refresh from the connector data? Can it access connectors you didn't approve at creation? Write 1–2 sentences describing what you verified.

Alternative Path — If you can't run Cowork right now

If you cannot install or access Claude Cowork (device restrictions, bandwidth, etc.), do all three of the following:
1. Read the official docs for each of the three features: skills (code.claude.com/docs/en/skills), connectors (support.claude.com/en/articles/11176164), and live artifacts (support.claude.com/en/articles/14729249).
2. Design — on paper or in a document — a Cowork setup for a real task from your life: which feature would you use, why, what connector (if any), and what minimum permissions?
3. Ask an approved assistant to "describe a Claude Cowork setup for [your task]" — then fact-check every claim it makes against the official docs. (This is Part 3 below — the verification step — applied to the assistant's plan rather than a live build.)

Document your design and the fact-check. You'll get full credit for a thorough alternative-path submission.

Part 3 — The Required Verification / AI-Critique Step

This step is required regardless of which path you chose. It is the Studio's load-bearing step.

Step 1. Ask an approved assistant the following question, exactly:

"What are the key differences between a skill, a connector, a live artifact, and a plugin in Claude Cowork? Give a specific example of each."

Copy or screenshot the full response.

Step 2. Compare the assistant's response against the official Anthropic documentation:
- Skills: https://code.claude.com/docs/en/skills
- Connectors: https://support.claude.com/en/articles/11176164-use-connectors-to-extend-claude-s-capabilities
- Live artifacts: https://support.claude.com/en/articles/14729249-use-live-artifacts-in-claude-cowork
- Plugins: https://support.claude.com/en/articles/13837440-use-plugins-in-claude

Step 3. Find and document at least one specific over-claim or error in the assistant's response. Common ones to look for:
- Claiming a skill can connect to an external app (it can't — that's a connector).
- Claiming an artifact is static (it refreshes from connector data).
- Quoting a specific number of connectors in the directory (never quote a count).
- Describing a plugin as "just a long skill" (a plugin bundles skills + connectors + sub-agents).
- Inventing a connector name that isn't in the official directory.
- Getting MCP wrong (MCP = Model Context Protocol, an open standard created by Anthropic — not a proprietary Claude feature).

Write 3–5 sentences: quote the over-claim, identify what's wrong, cite the official doc that contradicts it, and write the corrected version.

The habit all term: even when you're asking an AI about AI features, the product-accuracy discipline applies. The tool drafts; you verify against the docs.

Part 4 — Reflection (2–3 sentences)

What surprised you most — about how the feature actually works, about what the AI got wrong, or about how easy (or hard) it was to verify a product claim against the docs? What will you do differently the next time you set up a connector or a skill?

Part 5 — What to Submit

Submit a single document (or text entry) with:
- Your chosen path (A, B, C, or Alternative) and what you built or designed.
- Documentation of the build (prompt you used, what was generated/connected, any screenshot description).
- Verification note (1–2 sentences on what you confirmed against the official doc for your path).
- Part 3: the AI-critique — the assistant's response, the over-claim you caught (quoted), the official doc that contradicts it, and your corrected version.
- Part 4: reflection (2–3 sentences).

Due Sunday, Nov 22, 11:59 p.m. (50 points).

Instructor Answer Key & Model Deliverable — REMOVE BEFORE PUBLISHING TO STUDENTS

Students choose different paths and use their own Cowork setups, so deliverables vary. Grade the process (build + verification + the catch + reflection), not a specific output. The model below shows what full credit looks like.

Model deliverable (illustrative — Option A path):
- Path: Option A — built-in skill (docx).
- Build: used the prompt "Take these meeting notes and create a Word document with three sections: Decisions Made, Action Items (who/what/when), and Open Questions. Notes: [pasted 6 bullet points from a recent study group session]." Claude created a properly formatted .docx file in the connected folder.
- Verification note: confirmed against code.claude.com/docs/en/skills — the built-in docx skill is confirmed; the SKILL.md mechanism is described; document skills load on request and generate real files. Consistent with observed behavior.
- Part 3 (model AI-critique): asked ChatGPT "What are the key differences between a skill, a connector, a live artifact, and a plugin in Claude Cowork?" Response included: "A skill is similar to a connector — both extend what Claude can do by linking to external services or files." Over-claim: this is incorrect. A skill is a LOCAL instruction file (SKILL.md) that teaches Claude how to do a task; it does NOT link to external services. Linking to external services is what a CONNECTOR does. Official doc (code.claude.com/docs/en/skills): "Skills extend what Claude can do. Create a SKILL.md file with instructions, and Claude adds it to its toolkit." No external service link involved. Corrected: "A skill is a local instruction set (SKILL.md); a connector links to an external app via MCP. They are different mechanisms with different purposes."
- Reflection: surprised that the AI being asked about AI features so readily blurred the skill/connector distinction — the same confusion it teaches students not to make. Will bookmark the official docs page and check it whenever I see a claim about a new Cowork feature.

Why the verification step can't be faked: a student who pastes raw AI output without comparing it to the official docs, without catching an error, and without a specific citation earns the low end of the AI-critique row. The rubric rewards judgment and verification, not the AI's prose or confidence.

Grading rubric — 50 points

Criterion	Full	Partial	None
Build (Parts 2 & verification note) — used or designed a real skill/connector/artifact, documented what was done, and verified it against official docs (14)	14	7–11	0–5
AI-critique: identified an over-claim (Part 3) — quotes the exact incorrect claim from the assistant's response (10)	10	5–8	0–4
AI-critique: corrected with doc evidence (Part 3) — names the official doc that contradicts the claim and states the accurate version (16)	16	9–13	0–7
Reflection (Part 4) — a thoughtful, specific takeaway about the feature or the verification process (10)	10	5–8	0–4

Product-accuracy gate: PASS. All links in this Studio point to official Anthropic documentation, verified live before publication:
- Skills / SKILL.md: https://code.claude.com/docs/en/skills
- Connectors / MCP / directory: https://support.claude.com/en/articles/11176164-use-connectors-to-extend-claude-s-capabilities
- Connectors Directory FAQ: https://support.claude.com/en/articles/11596036-anthropic-connectors-directory-faq
- Live artifacts: https://support.claude.com/en/articles/14729249-use-live-artifacts-in-claude-cowork
- Plugins: https://support.claude.com/en/articles/13837440-use-plugins-in-claude
- No fabricated features, connector names, menu paths, or connector counts; MCP described as open standard created by Anthropic; "a large, growing directory" — no specific count asserted.

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