← Back to the Using Artificial Intelligence outline The Course Maker
← Using Artificial Intelligence outline
Week 16 · Lecture outline

Week 16 — Lecture Outline · Final Review & Exam (+ Capstone)

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

Course: Using Artificial Intelligence (AI 101) · Silver Oak University (fictional sample) · Prof. Quinn
Objectives covered: cumulative — Objectives 1–8 (Weeks 1–15).
Obj 1 — what generative AI is, how it works, and its limits; Obj 2 — effective prompting (conversation, content, emphasis, meta, structured, examples, simulations); Obj 3 — modalities and choosing the right tool; Obj 4 — verification and critical thinking; Obj 5 — Claude Cowork: projects, files, skills, connectors, artifacts; Obj 6 — automation: scheduled tasks, dispatch, computer use, Chrome, Excel, safely; Obj 7 — responsible AI: privacy, ToS, IP, bias, integrity, ethics; Obj 8 — integration and the capstone.

Meeting pattern: 2 sessions × 75 min = 150 min. Segment minutes total ~150; scale to your section.

This is the final review-and-exam week — no new content. It is cumulative over the entire course (Weeks 1–15, Objectives 1–8). Each segment briskly re-teaches one or two objectives with its highest-yield ideas, the single hardest misconception, and a worked example where appropriate. The final segment frames the capstone and the Final exam. This week's only graded item is the Final (25%) — there is no quiz, no discussion, no assignment, and no AI Build Studio; the Final stands in for all of them. The Final pairs with a Study Guide (M) + Exam-Prep Tutorial (N) + Practice Final (O), built separately and referenced here by name. AI is not permitted on the Final.

Week at a Glance
The week's big question "Across all eight objectives — how AI works, how to prompt it, which tools to use, how to verify output, how to build and automate with Cowork, how to use AI responsibly, and how to put it all together — can I name the move each topic requires, catch the classic mistake, and demonstrate the embed-don't-trust discipline?"
By the end of the week, students can… Re-run each objective's core move on demand: (1) explain next-token prediction, the context window, and confident wrongness; (2) fix a weak prompt, apply all prompting skills, counter sycophancy; (3) match tool to task, execute the record→transcribe→analyze workflow; (4) run the verification workflow and explain why AI confidence ≠ AI accuracy; (5) distinguish agent from chatbot, define skill/connector/artifact/plugin, name the MCP open standard; (6) state the scheduled-task constraint (awake + app open), describe dispatch, match computer use/Chrome/Excel to their surfaces, name the money rule, identify prompt injection; (7) apply the billboard test, classify data as safe-or-not, state IP basics with the "not legal advice" caveat, explain AI bias; (8) design, build, verify, and reflect on an end-to-end AI workflow. Also: frame the capstone and walk into the Final knowing its coverage, weight, and prep strategy.
Key vocabulary (all review) generative AI; LLM; next-token prediction; training; context window; hallucination; sycophancy; confidently wrong; Turing test; conversation; providing content; emphasis (Markdown, XML, CAPS); meta-prompting; structured-prompt components (Context/Role/Goal/Audience/Constraints/Voice-Format/Data-Logic/Examples/Evaluation); zero/one/few-shot; simulation; voice prompting; record→transcribe→analyze; image generation; modality; tool landscape (chatbots, image, audio/music, video, research, coding); verification workflow; citation hallucination; fabricated statistics; cross-checking; agent; chatbot; Claude Cowork; project; connected folder; task; skill; SKILL.md; connector; MCP (Model Context Protocol; open standard; Anthropic); live artifact; plugin; scheduled task; dispatch; computer use; Claude in Chrome; prompt injection; Claude in Excel; billboard test; HIPAA; FERPA; PCI; ToS; data retention; IP; copyright; bias; academic integrity; embed-don't-trust; capstone
Materials slides (Deck 16 — the final-review deck, 14 slides, blue-dominant), the Study Guide (M), the Exam-Prep Tutorial (N) (AI-based), the Practice Final (O), scratch paper, one approved chatbot (ChatGPT, Claude, or Gemini) for the verify-the-AI moment
Timing note 8 segments, ~150 min total. Session 1 (Tue) = Segments 1–4 (~75 min): map + Objectives 1–4 (how AI works → prompting → modalities → verification). Session 2 (Thu) = Segments 5–8 (~75 min): Objectives 5–8 (Cowork → automation → ethics → capstone + Final frame). Scale to your section.

Segment 1 — Hook & the Map of the Whole Course (10 min) · Session 1 opens

Hook. Write on the board with no comment: "An AI assistant that gives a confident, fluent answer is more likely to be accurate than one that hedges." Ask: "True or false?" Let the room react, then reveal: False — this is one of the core misconceptions the entire course exists to correct. Confidence in tone does not equal accuracy in fact. A model that says "Here are the three key factors…" in a polished paragraph can be fabricating all three. The skill this course built — embed-don't-trust — is the antidote.

The map (say it out loud as you display Slide 3):

WHAT AI IS AND HOW IT WORKS: Obj 1 (how LLMs work, context window, confident wrongness) · Obj 2 (the prompting arc: conversation → meta → structured → examples → simulations).
TOOLS AND VERIFICATION: Obj 3 (modalities and the tool landscape) · Obj 4 (hallucination shapes and the verification workflow).
COWORK — BUILD AND AUTOMATE: Obj 5 (projects, files, skills, connectors, artifacts) · Obj 6 (scheduled tasks, dispatch, computer use, Chrome, Excel, safely).
ETHICS AND INTEGRATION: Obj 7 (privacy, ToS, IP, bias, integrity, ethics) · Obj 8 (the capstone — design, build, verify, reflect).

Why it matters line (memory hook): "The whole course is one sentence — AI generates plausible text, not verified truth; so your job in every workflow is to use its speed and range while supplying the judgment, the verification, and the ethical guardrails it cannot provide itself. Embed, don't trust."

Segment 2 — Objectives 1 & 2 Review: How AI Works & the Prompting Arc (22 min)

Re-teach Obj 1 in plain language. A large language model predicts the next most plausible token based on statistical patterns in training data — it does not think, understand, or retrieve verified facts. The context window is finite; a very long conversation pushes earlier content out (why the AI seems to forget). AI is confidently wrong because it generates fluent text even when fabricating — confident tone does not mean accurate content. AI differs fundamentally from a search engine: a search engine indexes and retrieves; an LLM generates. The Turing test (Alan Turing, 1950, "Computing Machinery and Intelligence") tests whether a machine's text is indistinguishable from a human's — it does not prove understanding, consciousness, or factual reliability.

Re-teach Obj 2 in plain language. The prompting arc from Weeks 3–6:
- Skill 1 — have a conversation, ask for guidance; counter sycophancy (the AI agrees with you even when you are wrong) by explicitly asking for critique.
- Skill 2 — provide content (paste your notes, a document, data) with emphasis (Markdown bold/headers, XML-style tags <focus>...</focus>, CAPS for must-dos) to direct the AI's attention.
- Skill 3 — emphasis techniques: Markdown, XML tags, capitalization — not politeness, actual structural signals.
- Skill 4 — meta-prompting: ask the AI to help write or improve the prompt ("Ask me clarifying questions one at a time, then return a Markdown prompt").
- Skill 5 — the structured-prompt components: Context · Role · Goal · Audience · Constraints · Voice/Format · Data/Logic · Examples · Evaluation. The reusable template.
- Skill 6 — zero/one/few-shot examples: give the AI examples to teach it a voice, format, or task. (Few-shot = several examples — not exactly one.)
- Skill 7 — simulations: role-play a difficult conversation, a pre-mortem, a decision scenario, or an adaptive tutor. Classic trap: AI-generated quotes attributed to a simulated historical figure are generated text — never cite them as real.

One worked example (do it live): Show a weak prompt — "Write something about privacy." Then build it up, adding components one at a time: Role + Goal + Audience + Constraints + Format = a structured, effective prompt. Ask the class to identify what was added.

Highest-cost misconceptions + cures:
- ❌ "AI understands what I mean." → ✅ It predicts plausible next tokens — it does not understand.
- ❌ "Sycophancy means the AI is smart." → ✅ Sycophancy means the AI will agree with you even if you're wrong — always ask for critique explicitly.
- ❌ "Few-shot means exactly one example." → ✅ Few-shot = several examples; one example = one-shot.
- ❌ "Quotes a simulated historical figure says are real." → ✅ They are generated — never cite them as the person's actual words.

Segment 3 — Objectives 3 & 4 Review: Modalities, Tools & Verification (20 min)

Re-teach Obj 3 in plain language. AI works across multiple modalities beyond text: voice prompting (speak instead of type); the record → transcribe → analyze workflow (record a meeting with a free recorder, transcribe with a transcription tool, then ask an AI to produce minutes and action items — checking for transcription errors and summary fabrications); image-to-text, handwriting recognition, and document analysis; image generation (DALL·E, Midjourney, Adobe Firefly). The tool landscape: text chatbots (ChatGPT, Claude, Gemini, Copilot, Grok, Meta AI); image (DALL·E, Midjourney, Adobe Firefly); audio/music (Suno, Udio, ElevenLabs); video (Sora, Google Veo, Runway); research (NotebookLM, Perplexity); coding (Claude Code, GitHub Copilot, Cursor). Match the tool to the job — no single chatbot does everything equally well.

Re-teach Obj 4 in plain language. Hallucination shapes: invented citations (the source does not exist), fabricated statistics (the number is made up), fake case law, wrong arithmetic, fabricated quotes from real people. Sycophancy: the AI agrees with the user even when wrong. The verification workflow: (1) ask for sources and check them in a library database or authoritative site; (2) cross-check in a second model or independent source; (3) ask the AI to critique itself — "where might this be wrong?"; (4) request hedging — "if you're not certain, say so." Critical rules: asking the same model to verify itself is not reliable; two AIs agreeing is not proof.

One worked example (verification): Write a plausible-looking but invented citation on the board. Walk through the four verification steps and show what you would find (or not find) in a library database.

Highest-cost misconceptions + cures:
- ❌ "Chatbots are text-only." → ✅ Modern chatbots handle voice, images, and documents.
- ❌ "If the AI provides a citation, the source is real." → ✅ Citations are one of the most common hallucination shapes — always check.
- ❌ "Asking the AI to check itself fully verifies it." → ✅ The model has the same biases that produced the error; cross-check externally.

Segment 4 — Break + Quick Interaction (3 min)

Quick interaction (think-pair-share): "An AI gives you a beautifully formatted 500-word essay on a topic you asked about. It contains three citations. What are the first two things you do?" (Answer: verify at least one citation in an external source; check whether the essay actually addresses the goal you set — don't just trust that it does because it looks polished.) This reinforces both Obj 2 (was the prompt clear?) and Obj 4 (verify before you trust). Transition to Session 2 / Segment 5.

Segment 5 — Objective 5 Review: Cowork — Projects, Files, Skills, Connectors, Artifacts (15 min) · Session 2 opens

Re-teach Obj 5 in plain language. The vocabulary table — lock these definitions (they will appear on the Final as a matching item):

Term Definition
Agent Software that takes multi-step actions on your behalf — reads files, writes outputs, executes tasks — as opposed to a chatbot, which replies in one turn
Project A persistent, self-contained Cowork workspace with its own files, instructions, and memory that carries context across multiple tasks
Connected folder A local folder on your computer that Claude can read files from and write files to
Skill A reusable instruction set stored as a SKILL.md file that teaches Claude how to perform a specific type of task; built-in skills include docx, pptx, xlsx, pdf
Connector (MCP) A link to an external app using MCP — Model Context Protocol, an open standard created by Anthropic; each connector runs only with the permissions you grant
Live artifact A persistent, interactive view that refreshes with current data from your connected apps each time you open it (not a static file)
Plugin A bundle of skills + connectors + sub-agents packaged as one installable unit

Core confusions to cure:
- ❌ "A skill and a connector are the same thing." → ✅ A skill is a local instruction set (SKILL.md); a connector is an MCP link to an external app. To pull Gmail data, you need the Gmail connector — a skill cannot connect to an external app.
- ❌ "An agent is just a chatbot." → ✅ An agent takes multi-step actions; a chatbot replies in one turn.
- ❌ "A live artifact is a static file." → ✅ A live artifact refreshes from your connected data every time you open it.
- ❌ "MCP is proprietary to Anthropic." → ✅ MCP is an open standard — other AI systems can implement it too.

Source: Get started with Claude Cowork · Use connectors · Extend Claude with skills

Segment 6 — Objective 6 Review: Automation — Scheduled Tasks, Dispatch, Cross-App Workflows, Safety (18 min)

Re-teach Obj 6 in plain language. The automation vocabulary and rules:

Scheduled tasks: set work to run automatically (recurring or one-time). Type /schedule in a Cowork task, or use the Scheduled area in the sidebar. The critical constraint: scheduled tasks run ONLY while your computer is awake AND the Claude desktop app is open. They do not run on Anthropic's servers. If the computer is asleep when the task is scheduled, Cowork skips the run and re-runs it when the machine wakes up. Source: Schedule recurring tasks in Claude Cowork.

Dispatch: kick off work to run autonomously in the background, then receive a notification with the result. Different from a regular interactive task — asynchronous rather than synchronous. Source: Assign tasks from anywhere in Claude Cowork.

Computer use: Claude controls native desktop applications (screenshots, clicks, keyboard). Claude in Chrome: Claude navigates and interacts with Chrome browser tabs — carries prompt-injection risk (malicious instructions hidden in web page content can redirect Claude's actions; defend with approval checkpoints, start on trusted sites, watch for unexpected behavior). Available in beta on paid plans. Source: Claude for Chrome. Claude in Excel: a sidebar inside Microsoft Excel where Claude can read, analyze, modify, and create workbooks. Available in beta on paid plans.

The absolute safety rules:
1. You — not any AI agent — execute financial transactions, trades, and purchases. Never automate money movement.
2. Approve before irreversible actions. Submitting a form, sending an email, deleting a file — add an approval checkpoint before each.
3. Least privilege. Connect only the folder, connector, or permission that the task actually needs — no more.
4. Link safety. Treat links in emails or messages as suspicious; verify the real URL before following.

Core misconception to cure:
- ❌ "Scheduled tasks run even if my computer is off." → ✅ They run ONLY while the computer is awake AND the Claude app is open.
- ❌ "Computer use and a connector are the same thing." → ✅ Computer use provides visual control of any native desktop app; a connector is an API-backed link to a specific service.

Verify-the-AI / Technology Workflow moment: Pull up a free chatbot and ask it to describe what happens to a Cowork scheduled task when the computer is closed. Common AI over-promise: "It runs on Anthropic's cloud servers." Show the class why this is wrong and cite the official doc. This is the discipline the exam will test.

Segment 7 — Objectives 7 & 8 Review: Ethics, Privacy & the Capstone (17 min)

Re-teach Obj 7 in plain language. The ethics and privacy framework from W15:

  • Billboard test: before pasting content into a free AI tool, ask — "Would I be okay if this became publicly visible?" If no, don't paste it.
  • Data types to never paste: HIPAA-protected health information (patient diagnoses, treatment notes); FERPA-protected student records (grades, disciplinary records); PCI payment card data (credit card numbers); employer confidential or proprietary material (unreleased product roadmaps, client contracts).
  • ToS basics: most consumer AI tools may retain inputs and use them for model improvement unless you opt out. Enterprise/paid plans generally have stronger protections.
  • IP and copyright: the U.S. Copyright Office has generally required meaningful human authorship for copyright protection; purely AI-generated work is in a contested, evolving legal status. This is not legal advice — consult an attorney for commercial use.
  • Bias: AI models are trained on human-generated data that reflects existing societal biases. More data does not mean more neutral. Review AI outputs for fairness, especially when they represent or affect people.
  • Academic integrity: submitting AI-generated work as your own when original work is required is academic dishonesty. AI is not like a calculator — it can generate the entire assignment.
  • Skill 13 — troubleshooting: if AI gives confused or contradictory answers in a long conversation, start a fresh conversation (context-window overload is the most common cause).

Re-teach Obj 8 — the capstone synthesis. Obj 8 asks you to put everything together. The capstone you designed, built, verified, and reflected on is the practical demonstration. The four required elements were: the design (what problem does it solve, what tools does it use), the build (it must actually run), the verification (document at least one error and how you fixed it), and the ethics reflection (what data did you connect, what did you choose not to automate). The Final will test Obj 8 with scenario items: given a described workflow, identify the error, name the fix, and apply the relevant safety or privacy principle.

Highest-cost misconceptions + cures:
- ❌ "Pasting client health data into a free tool is fine if the AI promises privacy." → ✅ HIPAA obligations apply to you, not to what the AI promises — never paste.
- ❌ "AI-generated content is automatically mine to copyright." → ✅ Contested and evolving; meaningful human authorship is currently required; not legal advice.
- ❌ "AI is neutral because it was trained on a lot of data." → ✅ The training data reflects human biases — more data does not mean unbiased.
- ❌ "Automations never need review after the first run." → ✅ Automated outputs inherit AI's failure modes; review the first several runs and set up ongoing spot-checks.

Segment 8 — Capstone Frame + Final Logistics + Callback & Close (5 min)

The capstone (re-state the four elements, briefly): design a real workflow → build and run it → verify (catch and fix at least one error) → ethics reflection (what you connected, what you chose not to automate, your safe-use posture). Assessed under the Final. Document it and submit the documentation before the Final closes.

Callback to the hook: "We started today with a false statement — 'confident, fluent AI output is probably accurate.' Now you have the full eight-objective toolkit to know exactly why that is false and exactly what to do instead. Embed the verification step. Don't trust the fluent tone. Use your own judgment. That is the whole course."

Final logistics:
- 25 items, 4 points each, 100 points total. Mixed types: multiple-choice, matching (term→definition + tool→best-use), true/false. No free-text entry.
- Coverage (proportional to teaching time): Obj 1 ≈ 3 items · Obj 2 ≈ 3 items · Obj 3 ≈ 2 items · Obj 4 ≈ 2 items · Obj 5 ≈ 4 items · Obj 6 ≈ 4 items · Obj 7 ≈ 4 items · Obj 8 ≈ 3 items.
- Window: opens Mon Dec 14; due six days later.
- AI is not permitted on the Final.
- Prep kit: Study Guide (M) → Exam-Prep Tutorial (N, submit share link) → Practice Final (O, ungraded).

Close: "You have built things, caught mistakes, debugged automations, and written ethical frameworks. What you have is not just AI literacy — it is judgment. Take it with you. See you at the Final. Good luck."

Instructor FAQ

Question Answer
"Do I need to memorize the MCP doc links for the exam?" No. Know what MCP is (Model Context Protocol, open standard, Anthropic), what connectors do, and the doc-verified Cowork behaviors — you won't be asked to reproduce URLs.
"Is the capstone graded separately?" No — it is assessed under the Final, not as a separate grade category. The grading map stays at 100%.
"Can I use AI to build my capstone?" Yes — AI use is required on all coursework including the capstone. The capstone must run, and you must verify and reflect on it.
"What if I can't install Cowork?" The capstone can be completed with any combination of AI tools from the course. Cowork is preferred for Objectives 5–6 integration but not strictly required.
"Will the Final have exactly the same format as the Practice Final?" Same blueprint and item types; different items (the practice shares no items with the live exam).

Scope Flag

This outline is the final cumulative review — it covers all 15 prior weeks in condensed form. Every concept, tool name, and Cowork feature claim in this outline was verified against the official Anthropic documentation (support.claude.com / docs.claude.com / code.claude.com) and the VERIFIED_FACTS.md reference as of 2026-06-29. No product feature, menu path, plan tier, or constraint is asserted without a documentation source.

Product-accuracy gate: PASS.

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