← Back to the Using Artificial Intelligence outline The Course Maker
← Using Artificial Intelligence outline
Week 5 · Lecture outline

Week 5 — Lecture Outline · Prompting III: Examples, Structure & Control

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 covered: Objective 2 — Use effective prompting techniques — including conversation, content, emphasis, meta-prompting, structured prompts, examples, and simulations — to produce high-quality, well-verified AI output.
SLOs touched: A (produce high-quality results with AI through strong prompting) · B (evaluate and use AI critically)
Meeting pattern: 2 sessions × 75 min = 150 min. Segment minutes below total ~150; scale to your own pattern.

Week at a Glance
The week's big question "When should you show the AI what you want — and how do you keep it from over-generalizing from your examples?"
By the end of the week, students can… (1) define and apply zero-shot, one-shot, and few-shot prompting and explain that few-shot means a few examples, not exactly one; (2) use examples to teach voice, format, and PII-safe transformations; (3) apply the full control toolkit (count, structure, constraints, expansion, regenerate, sources, guidance); (4) explain why regenerating does not fix fabricated facts; (5) request sources and verify every link before trusting it.
Key vocabulary zero-shot, one-shot, few-shot, in-context learning, voice/format example, PII (personally identifiable information), placeholder, constraint, regenerate, expansion, requesting sources, asking for guidance, drift/over-generalization
Materials slides (Deck 5), the week's readings + video links, one approved assistant for the live demos and the tutorial
Timing note 8 segments, ~150 min total. Session 1 = Segments 1–4 (~75). Session 2 = Segments 5–8 (~75).

Segment 1 — Hook & the Problem (8 min) · Session 1 opens

Hook. Put this situation on the board: "You've tried four different prompts. Each time, you've described what you want in detail. The AI's output is almost right — but the voice is off, or the format drifts, or it keeps doing this one thing you didn't ask for. What's missing?" Have the room call out answers. Write them. Then: "Almost always, the answer is an example. Today we learn to show instead of tell."

Why it matters (write it on the board): "Describing what you want is good. Showing what you want is almost always better."

The promise for the week: "By Friday you'll have a technique for pinning down voice, format, and structure that beats any description — and a control toolkit that makes every other prompting technique you know sharper."

Where we are in the arc: Week 3 = conversation, content, emphasis. Week 4 = meta-prompting, structured components, reusable templates. Week 5 = examples, structure, and control — the final piece of Objective 2. After this, we move to simulations (Week 6) and then multimodal AI (Week 7).

Segment 2 — The Shot Vocabulary: Zero, One, Few (20 min)

Plain language first. When you give an AI a task, you're always making a choice — whether to include examples of the kind of output you want. This choice has a name: the number of "shots."

Put this on a slide and drill it explicitly:
- Zero-shot — you give the task with no examples. "Summarize this article in three sentences." The AI does its best from the instruction alone. Works fine for well-understood, generic tasks.
- One-shot — you give exactly one example before the task. "Here's a summary I wrote for another article [example]. Write one like that for this article [article]." One sample sets a pattern.
- Few-shot — you give two to five examples before the task. The AI infers the pattern from all of them, which makes it more likely to generalize correctly rather than over-fitting to a single sample.

The classic confusion — say it out loud: "Few-shot means a FEW examples — typically two to five. One example is one-shot, not few-shot. They are not the same technique. This will be on the quiz."

Why few-shot works — plain language: The model learns the pattern "in context" — not by retraining, but by observing the examples within the current conversation window. It's like showing a colleague three examples of how you format meeting notes before asking them to format the next one. The more consistent examples you show, the more precisely they capture what you want.

When to use which:
- Zero-shot: simple, generic, well-understood tasks ("translate this phrase").
- One-shot: when you have a clear template to show.
- Few-shot: when you need to capture voice, tone, rhythm, or a precise format that's hard to describe — especially when you've already tried to describe it and the output still drifts.

Memory hook: "Shots = examples. Zero, one, few. Few means several — not one."

Segment 3 — Examples for Voice, Format & PII-Safe Use (22 min)

Examples for voice. Paste this scenario on the board (or live, on an assistant):

A student has three examples of her own social-media captions — casual, punchy, ends with a question. She pastes them and asks: "Write two more in exactly this voice for [new topic]."

Do it live. Paste three sample captions (write them for the demo — any short, casual, question-ending captions will do). Then ask the AI to continue the pattern. Compare the output: does it capture the rhythm and the question-ending habit? Where does it drift?

The key insight: you gave it voice through demonstration, not description. Describing "casual, punchy, ends with a question" is less precise than showing three examples of it.

Examples for format. Same technique, applied to structure:

You want a competitor comparison table with specific columns: Product, Price Range, Best For, Key Weakness. You paste one filled-in sample row and say: "Fill in the next five products the same way."

The AI infers the column order, the level of detail per cell, and the vocabulary level from your sample — without a lengthy format description.

PII and examples — a critical caution:

When you use real documents (an email, a customer record, a student submission) as few-shot examples, you must scrub sensitive information before pasting. The technique:
1. Replace real names with [NAME].
2. Replace dates with [DATE].
3. Replace account numbers, IDs, phone numbers, addresses with [ID], [PHONE], etc.
4. Paste the scrubbed version as the example.
5. Ask for the transformation; restore real values afterward if needed.

Memory hook: "Before you paste a real document, scrub the sensitive parts. Placeholders protect people."

Misconception + cure:
- ❌ "I'll just let the AI detect and remove the PII automatically."
✅ Cure: the AI may miss PII, infer it from context, or produce output that still contains it. You scrub before pasting — that's the only safe approach.

Segment 4 — Misconceptions + Quick Interaction (25 min) · Session 1 closes (~75)

Name the misconceptions out loud, then cure each:

  • ❌ "Few-shot means exactly one example."
    ✅ Cure: one example = one-shot. Few-shot means a few — typically two to five. This is not just vocabulary pedantry: with only one example the AI can over-fit to that single sample; more examples help it generalize.

  • ❌ "Asking for sources guarantees real ones."
    ✅ Cure: an AI can produce perfectly formatted, completely plausible-sounding citations that do not exist. Requesting sources is a useful move; verifying every link or citation is the mandatory next step.

  • ❌ "Regenerating fixes the facts."
    ✅ Cure: regenerating produces a different output — which may be just as wrong. A new set of invented citations is still invented. Regenerating is for stylistic variety, not accuracy correction. Fix bad facts by verifying and correcting explicitly.

  • ❌ "More examples are always better."
    ✅ Cure: too many examples can cause the AI to over-fit to surface features in your examples rather than the underlying pattern you want. Two to five is typically the sweet spot. Beyond that, explicit constraints may be more precise.

Interaction — Classify the Shot (rapid-fire, ~10 min):
Put four prompt scenarios on a slide; for each, students call out: zero-, one-, or few-shot?
- "Translate this sentence into Spanish." → zero-shot.
- "Here's a haiku I wrote [example]. Write one more for this topic." → one-shot.
- "Here are three of my past email intros [examples]. Write an intro for this new email." → few-shot.
- "Write me a cover letter for a marketing job." → zero-shot.

Then: "Why does it matter which one you're doing? Because the choice of how many examples to include changes how precisely you're communicating what you want — and how much you're risking over-generalization."

Segment 5 — The Control Toolkit (20 min) · Session 2 opens

Hook back in: "Last session: showing the AI what you want. Today: the full toolkit for controlling how much, what shape, and what quality it gives you."

Walk through six moves with a live example for each:

1. Specify a count. "Give me exactly five bullet points" beats "give me some bullet points." The AI interprets "some," "a few," and "several" loosely. Exact numbers work better. Demo: prompt with "some" → prompt with "exactly 5."

2. Request structure. Describe the precise output format. "Return a Markdown table with three columns: Term, Definition, Example." Or: "Give me a numbered list, not bullets." The more specific the format request, the less the AI interprets.

3. Set a constraint. Negative or positive limits narrow the output. "No jargon" / "each bullet under 15 words" / "include exactly one concrete example per point" / "do not include costs — I'll add those." Constraints prevent drift.

4. Ask for expansion. After getting an outline or a list, zoom into one item: "Take bullet 3 and write a full paragraph, including a real-world example." This is different from regenerating — you're deepening one specific part, not replacing the whole output.

5. Regenerate. Ask for a new version of the same output. Useful for stylistic variety or when the first attempt was a poor draw. Critical: does NOT fix factual errors. A regenerated set of fabricated citations is still fabricated.

6. Ask for guidance. "What information would help you give me a better answer?" / "What are the key tradeoffs here I should weigh?" / "What am I not thinking about?" Treats the AI as an advisor. The output still requires your judgment — but it can surface angles you hadn't considered.

Memory hook: "Count, structure, constrain, expand, regenerate, guide. Six moves, one goal: the exact output you actually need."

Segment 6 — Requesting Sources (Live Demo with Verification) (20 min)

Set up the scenario: "One of the most common things students ask AI to do is provide evidence — articles, statistics, citations. And it will. Let's look at exactly what you get, and what you have to do next."

Do it live (narrate every step):
1. Ask an approved assistant: "Give me five peer-reviewed studies on how sleep affects academic performance, with authors, journal names, publication years, and DOIs."
2. Read the output aloud. The citations look real — authors, journals, years, DOIs.
3. Open one or two in a new tab. (Live demo: at least one will not resolve — a 404, a wrong DOI, or an article that doesn't exist at the journal's site.)
4. Say aloud: "This is what fabricated citations look like. They're not random nonsense — they're plausible enough to use by accident."

The verification workflow:
1. Request sources.
2. Open every link in a new browser tab before trusting.
3. Confirm the article title matches what appears on the journal/publisher site.
4. If a DOI leads to a 404 or a different paper — it's fabricated.
5. Cross-check with Google Scholar or your library database.
6. If in doubt, ask the AI: "Are you certain this citation is real? If not, say so." (It may admit uncertainty — and that admission is useful.)

Land the key idea: requesting sources is a useful starting point for finding search terms and topic areas — it is not a substitute for verification. Never cite an AI-provided reference without independently confirming it exists.

Misconception + cure:
- ❌ "If the AI gave it a DOI and a journal name, it must be real."
✅ Cure: DOIs and journal names can be fabricated. The only proof is opening the link and seeing the paper at its real location.

Segment 7 — Technology Workflow with a Verify-the-AI Moment (18 min)

The complete few-shot workflow, step by step:
1. Choose your task and decide whether description alone is enough, or you need examples.
2. Write two to five examples of the output you want (for voice or format tasks).
3. Scrub any PII from real-document examples before pasting.
4. Specify control parameters in the same prompt: count, format, constraints.
5. Send and read the output — specifically watch for drift or over-generalization from your examples.
6. Iterate — if it drifts, add another example that shows the range, or add an explicit constraint that excludes the over-applied element.
7. Verify — if any output includes specific facts, dates, names, or citations, check them independently.

AI-critique / verify-the-AI moment:

Ask an approved assistant: "Give me three examples of a clean, punchy one-sentence description of what a large language model is — for a non-technical college student." Then: "Now give me the same in a more formal, academic voice." Read both outputs. Watch whether the "formal" version is genuinely more precise, or just adds jargon. Does "formal" mean hedged and passive — or does it mean clear and citable? That judgment is yours to make. The AI doesn't know what "formal" means to your audience — only you do.

Segment 8 — Callback, Tease & Hand-off (17 min) · Session 2 closes (~75)

Callback + tease:
- Callback: "Three weeks of prompting — now complete. You started with conversation and content (Week 3), moved to meta-prompting and structured components (Week 4), and this week added examples, control, and source-verification. That's the full Objective 2 toolkit."
- Tease next week: "Week 6 takes everything you know about prompting and uses it in a more ambitious way: simulations. You'll design a multi-turn role-play — a difficult customer, a project pre-mortem, a historical figure interview. But there's a critical caution that comes with historical-figure simulations: the AI's 'words' are generated, not real. Never present a simulated quote as a real one. We'll dig into that directly."

Hand-off (the week's graded work):
- Lecture Tutorial 5 — few-shot vocabulary + control techniques, AI tutor, share-link submission.
- Quiz 5 (no AI) — zero/one/few-shot, control moves, requesting sources, regenerate ≠ fix.
- Discussion 5 — "Voice, Authenticity & Demanding Sources."
- Assignment 5 — "Examples, Control & Verification."
- AI Build Studio 5 — "Few-Shot Your Format/Voice" — give the AI examples, control the output, catch the drift.

Instructor FAQ — Common Stumbles

Student says / does Quick cure
Calls any example-based prompt "few-shot." "One example = one-shot. Few-shot requires more than one — typically two to five."
Thinks regenerating fixes hallucinated citations. "Regenerating gives a different output — the facts are just as likely to be wrong. Verification is always manual."
Asks the AI for sources, trusts the output. "Open every link before using any citation. DOIs and journal names can be fabricated — plausibly."
Pastes a real email as a few-shot example. "Scrub the PII first — replace names, dates, IDs with placeholders before pasting."
Uses only one example and says it's "few-shot." "One example = one-shot. You need at least two for few-shot. The difference matters for how well the AI generalizes the pattern."
Thinks more examples always help. "Two to five is usually the sweet spot; beyond that, explicit constraints are often more precise than adding more examples."
"The AI captured my voice perfectly — I can just publish this." "Check it carefully: did it over-generalize a quirk into a formula? Does it sound like you at your best, or like a caricature?"

Scope flag

This outline covers Skill 6 (zero/one/few-shot; examples for voice, format, PII-removal) and the full control toolkit (count, structure, constraints, expansion, regenerate, requesting sources, guidance) within Objective 2 (effective prompting). The legal/copyright dimensions of voice-mimicry are previewed in Discussion 5 but taught in depth in Week 15 (ethics/privacy/IP). Multimodal input (pasting images or documents) was introduced in Week 3; few-shot examples are a text-based technique and are kept distinct from multimodal input here. All tools named (ChatGPT/Claude/Gemini/Copilot) are factual; the instructor and institution are fictional. The buyer-facing verb for the product is "generates."

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