Week 5 — AI Build Studio · "Few-Shot Your Format/Voice"

Course: Using Artificial Intelligence (AI 101) · Silver Oak University (fictional sample) · Prof. Quinn
Objective: Objective 2 — use few-shot examples to control voice, format, and output structure; catch drift and over-generalization · SLO A (produce a quality result with AI) · SLO B (verify and improve AI output)
Worth 50 points · AI Build Studios group = 15% of the grade · Studio 5
Format: a hands-on build — you'll choose a real recurring task, use two or three examples to teach the AI your format or voice, apply control moves, and then catch where it drifts or over-generalizes from your examples.

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.

Part 1 — The Build Goal

By the end of this Studio you'll have produced three things for a real, recurring writing or formatting task from your own life or major:
1. A few-shot prompt — two or three examples of the output you want, plus control parameters (count, constraints, format spec).
2. The AI's controlled output — the new content it produced by learning from your examples.
3. A short drift/verification write-up — naming at least one place where the AI over-generalized from your examples, drifted, or got something wrong — and how you fixed it.

This is the week's skill in practice: show the AI what you want, control the output tightly, then catch where it over-applies the pattern.

Open one approved assistant to build in: ChatGPT (https://chatgpt.com), Claude (https://claude.com), Gemini (https://gemini.google.com), or Copilot (https://copilot.microsoft.com). A free account is enough.

Part 2 — Pick a Real Recurring Task

Choose something you actually do more than once — this makes the result useful across your life, not just a one-off. The task should have a consistent format or voice that you want to teach the AI. Examples:

• Writing intro sentences for assignments, reports, or emails in your major.
• Drafting social-media captions for a project, club, or personal account.
• Writing professional email subject lines or meeting-update bullet points.
• Formatting product descriptions, lab report sections, or club newsletter blurbs.
• Rewriting dense notes into a specific study-card format.

Avoid tasks with no consistent format (e.g., "write me an essay on anything") — few-shot only helps when there's a real pattern to teach.

Write your task here: I want AI to help me produce __ in the style/format of ____.

Part 3 — Build Your Few-Shot Prompt (two or three examples + control)

Step 1 — Write your examples. Write two or three examples of the output you actually want. These can be real past outputs you've written, or clean versions you write now that demonstrate the pattern. Requirements:
- Each example should show the same format and voice.
- They should be different enough from each other that the AI sees the pattern — not just a single quirk.
- Scrub any personally identifiable information (PII) before pasting: replace names with [NAME], dates with [DATE], IDs with [ID], etc.

Step 2 — Add control parameters in the same prompt. In addition to your examples, specify:
- A count: "give me exactly three more" (not "give me some").
- A structural constraint: one rule the examples alone don't guarantee — e.g., "each item under 20 words," "no exclamation points," "always start with a verb," "never use passive voice."
- An optional format request: "return as a numbered list" / "one per line" / "in a Markdown table."

Step 3 — Send and read. Paste your examples + control parameters in one prompt. Read the output carefully.

Your few-shot prompt goes here: (paste what you sent the AI — examples + count + constraint + format spec)

Part 4 — The Drift-Catch / AI-Critique Step (required — this is the BYOAI step)

Now audit the AI's output for drift or over-generalization. After the AI produces new content from your examples, it may:
- Mechanically repeat a surface feature that appeared in all your examples but wasn't meant as a rule (e.g., every example happened to use the word "powerful" — now every output does too).
- Lock in a sentence structure from the examples and apply it even when it doesn't fit.
- Over-constrain itself — e.g., if your examples all happened to be short, it might truncate output that should be longer.
- Miss the spirit of the voice while matching the surface: grammatically similar but tonally off.

Your required steps:
1. Read the output specifically for drift — not just "does it sound okay?" but "is the AI applying a pattern it shouldn't?"
2. Find at least one drift or over-generalization. If you can't find one, re-read more carefully — add a fourth output and watch if a pattern emerges.
3. Fix it using one of two moves:
- Add another example that shows the element is not always present (demonstrates range).
- Add an explicit constraint that excludes the over-applied element ("vary the opening; do not always use the same first word").
4. Also watch for factual accuracy if your output includes any specific claims, names, statistics, or links — verify these independently. AI can hallucinate even inside a few-shot prompt.

Write 3–4 sentences reporting: what drift or over-generalization you caught, what specifically the AI over-applied, and how you fixed it. If you also caught a factual error, report that separately.

The habit all term: the tool drafts; you judge. A chatbot can over-generalize from your examples and produce something that's technically consistent with them but wrong for the task. Catching that is the skill.

Part 5 — Reflection (2–3 sentences)

What surprised you most — about how well (or how poorly) the AI learned from your examples? What did you learn about the difference between describing what you want and showing what you want? What will you do differently the next time you use few-shot prompting?

Part 6 — What to Submit

Submit a single document (or text entry) with:
- Your task and the recurring situation it's for.
- Your few-shot prompt (the examples + control parameters you sent).
- The AI's output (the new content it produced).
- Your Part 4 drift-catch write-up (what you caught + how you fixed it).
- Your Part 5 reflection.

Due Sunday, Oct 4, 11:59 p.m. (50 points).

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

Students use their own tasks, so deliverables vary. Grade the process (examples quality → control → drift-catch → fix → reflection), not a specific answer. The model below shows what full credit looks like.

Model deliverable (illustrative):
- Task: "Write professional email subject lines for project update emails at my internship."
- Examples provided:
- "Project Alpha: testing complete, launch ready for Thursday"
- "Q3 budget update: final numbers attached"
- "Client feedback received: three changes needed by Friday"
- Control parameters added: "Give me exactly four subject lines for a project status update about a delayed software release. Each under 12 words. No exclamation points. Start each with the project name."
- AI output (example): four subject lines, all under 12 words, no exclamation points, starting with a project name. But: they ALL opened with "Project [Name]:" — every single one, even when other openings would be more natural.
- Part 4 (drift catch): "The AI locked onto the colon format and 'Project [Name]:' opening from all three examples and applied it to every output, even when a different structure would read better. My examples all happened to use this structure — but it wasn't a rule. I fixed it by adding a constraint: 'Vary the opening structure — not every line needs to start with the project name' and regenerated. The new output had more variety." (Also: "I checked whether 'launch ready for Thursday' was a real calendar claim — it was illustrative, not factual, so no verification issue here.")
- Reflection: notes that describing "professional and specific" would not have produced the same pattern as the examples, and plans to always add at least one example that shows the range — not just the center — of the pattern they want.

Why the drift-catch step can't be faked: a student who reports "AI matched my examples perfectly" without any evidence of reading critically for over-generalization earns the low end of the AI-critique row. The rubric rewards judgment — the ability to notice when the AI over-applied a pattern that was an accident of the training set.

Grading rubric — 50 points

Quality gate (self-checked): tools and links named (ChatGPT/Claude/Gemini/Copilot homepages) are real and current; no fabricated product features; the activity requires the student to catch the AI's over-generalization, not trust it (verification-as-content). No student-produced output is asserted as "the" answer — the key grades the process. Product-accuracy gate: PASS.