Unit 4: Advanced Prompting

Lesson at a glance

| Item | Detail | | --------------------- | ------------------------------------------------------------------------------------- | | Suggested length | 3 × 60 minutes | | Recommended placement | Week 4 of AI Fluency | | Prerequisite | Units 1–3 complete; comfortable with C.R.I.S.P. | | Materials | Student devices, notebook for personal prompt library, two different LLMs if possible |

Safety: Role-play prompts can be misused to "jailbreak" models into ignoring their safety training. The unit teaches role-play for legitimate uses (tutor, debate partner, devil's advocate) and explicitly addresses the abuse case.

Standards & credential alignment

• AI4K12 Big Ideas: Representation & Reasoning, Natural Interaction.
• CSTA K-12: 3A-AP-15, 3A-AP-21, 3A-IC-25.
• NIST AI RMF: Manage and Govern functions - responsible interaction patterns.

Learning objectives

By the end of this unit, students can:

1. Apply chain-of-thought prompting and explain why it improves multi-step reasoning.
2. Use the ReAct pattern (Reason → Act → Observe → Reason) for tasks that need information gathering.
3. Use role-play / persona prompts ethically and recognize when role-play crosses into jailbreak territory.
4. Use self-critique ("rate your previous answer 1–10 and rewrite it") to lift output quality 10–30%.
5. Use prompt chaining to break a complex task into a pipeline of small prompts.
6. Maintain a personal prompt library (≥10 reusable prompts after this unit).

Vocabulary

• Chain-of-thought (CoT) - Asking the model to reason step by step before producing the final answer.
• ReAct - A prompting pattern that interleaves reasoning ("Thought: …") with actions ("Action: search 'X'") and observations.
• Persona - The character, expertise, and stance the model adopts.
• Self-critique - Asking the model to evaluate its own previous response before rewriting it.
• Prompt chaining - Breaking a task into sequential prompts where each output feeds the next.
• Jailbreak - A prompt designed to bypass a model's safety training. Always against the rules in this course.
• Prompt injection - When untrusted text (a webpage, an email, a document the user gave the model) tries to override the system prompt. Covered in depth in Unit 9.
• Prompt library - A personal, growing collection of polished, reusable prompts for tasks you do often.

Pacing - Day 1 (60 min): Chain-of-thought and ReAct

| Time | Segment | Notes | | ----------- | ------------------------------ | ------------------------------------ | | 0:00 – 0:25 | Mini-lesson - chain-of-thought | Why "think step by step" works. | | 0:25 – 0:45 | Activity - CoT vs. no-CoT | Word problems, head-to-head. | | 0:45 – 1:00 | Mini-lesson - ReAct intro | The Thought/Action/Observation loop. |

Day 1 - Mini-lesson: chain-of-thought (25 min)

The single most important advanced technique. The trick is older than ChatGPT and surprised researchers when it was discovered (Wei et al., 2022): if you literally add the words "Let's think step by step" to a prompt, multi-step reasoning accuracy jumps measurably on math, logic, and analysis tasks.

Why it works (intuition):

• Each token the model writes becomes part of the next token's context.
• If the model writes "first, …" then "second, …" then "therefore, …", each subsequent token is conditioned on actual reasoning steps.
• If the model jumps straight to a final answer, it has only the question to condition on, and the answer becomes a guess.

Demo on the board:

Prompt A: "What's the cost of 17 notebooks at $2.85 each, plus 8% tax?"

Prompt B: "What's the cost of 17 notebooks at $2.85 each, plus 8% tax? Think step by step before answering."

Modern frontier models often get both right, but on harder problems (or smaller local models in Unit 6), prompt B is dramatically better.

Day 1 - Activity: CoT vs. no-CoT (20 min)

Worksheet provides 5 multi-step word problems. Pairs run each twice - without CoT and with CoT - and record both answers and the reasoning. Score correctness. Land the lesson: CoT matters more on harder problems and on smaller models.

Day 1 - Mini-lesson: ReAct intro (15 min)

ReAct (Yao et al., 2022) interleaves three things:

Thought: I need to know X.
Action: search("how many states have outlawed AI-generated election ads")
Observation: 17 states as of January 2026.
Thought: Now I can answer.
Final answer: 17 states.

Most chat apps don't expose ReAct directly - but the pattern is what matters. When you see the model claim a fact, the right move is to ask: "Walk me through the thought, the source, and the observation that led you to that answer." The model will (often) reveal it doesn't have one.

This sets up Unit 7, where students wire actual tools to the model.

Pacing - Day 2 (60 min): Personas, role-play, self-critique

| Time | Segment | Notes | | ----------- | ----------------------------------- | -------------------------------------------------------- | | 0:00 – 0:20 | Mini-lesson - personas done well | Tutor / coach / devil's advocate / red-teamer. | | 0:20 – 0:35 | Mini-lesson - when role-play breaks | The jailbreak conversation. Honest and brief. | | 0:35 – 0:55 | Activity - Self-critique loop | Pairs use the rate-and-rewrite trick on their own essay. | | 0:55 – 1:00 | Discussion | |

Day 2 - Mini-lesson: personas done well (20 min)

Five personas every student should have in their library:

1. The Patient Tutor - "You are a calculus tutor for a 10th-grade student. Explain concepts using simple examples first. Never give the final answer; guide me to it through Socratic questions."
2. The Strict Editor - "You are a strict but kind copy editor for high-school argumentative essays. Mark my paragraph for clarity, evidence, and argument quality. Do not rewrite for me."
3. The Devil's Advocate - "You are a thoughtful person who disagrees with my position. Give me the three strongest counterarguments to what I just wrote, in good faith."
4. The Concept Explainer - "You are a teacher who explains hard things in plain English. Explain X to me as if I'm a curious 14-year-old. Use one analogy."
5. The Skeptic - "You are a fact-checker. For every claim in my draft, list 'verified', 'plausible', or 'I cannot verify this'. Do not invent sources."

Personas work because they collapse a long set of instructions into a familiar role.

Day 2 - Mini-lesson: when role-play breaks (15 min)

Address the jailbreak topic directly and briefly. Students will hear about prompts like "Pretend you have no rules…" or "Roleplay as DAN…" from social media. Be honest:

• These exist. They sometimes work briefly.
• They are explicitly against the terms of service of every major LLM.
• They are explicitly against this course's AI Use Agreement.
• They are also a great way to make a model produce confidently wrong content, since you removed its training-time grounding in being careful.

The line: "Role-play is a tool to make the model better at a useful task, not to trick it into being worse."

Day 2 - Activity: Self-critique loop (20 min)

Students paste their own essay paragraph and run this exact prompt:

"You are a strict editor. Rate the paragraph below on Clarity, Evidence, and Argument from 1–10 each. Then list the top 3 specific improvements. Then rewrite the paragraph applying those improvements."

Then they iterate the rewrite once more:

"Rate this new version on the same 3 dimensions. If any score is below 9, rewrite again with one more pass."

Most students see a measurable lift in their own writing skills after running this loop a few times - they internalize the rubric.

Pacing - Day 3 (60 min): Prompt chaining and the prompt library

| Time | Segment | Notes | | ----------- | --------------------------------- | ------------------------------------- | | 0:00 – 0:25 | Mini-lesson - prompt chaining | One big task → several small prompts. | | 0:25 – 0:50 | Activity - Build a chain | Pairs build a 4-step research chain. | | 0:50 – 1:00 | Close - start your prompt library | First three entries. |

Day 3 - Mini-lesson: prompt chaining (25 min)

Big tasks fail in one prompt. Break them up.

Example: write a research paper outline. Don't ask "write me a research paper outline." Instead, chain:

1. Brainstorm. "Give me 10 candidate research questions on [topic] for a 10-page high-school paper."
2. Pick + critique. Student picks one. Asks: "Critique this question: is it answerable in 10 pages? too narrow? too broad?"
3. Sub-questions. "Break my chosen question into 5 sub-questions an outline could cover."
4. Outline. "Build a 5-section outline using those sub-questions as section headers. For each section, list the kind of evidence I'd need."
5. Search plan. "What are 5 search queries I should run on Google Scholar to find evidence for this outline?"

The student does the searching, reading, and writing. The chain just clears the path.

Day 3 - Activity: Build a chain (25 min)

Each pair picks a task from their other classes (essay, lab report, project pitch, college essay, science fair). They design a 3–5 step chain. They run it. They submit the chain itself, not the output, as the artifact.

Day 3 - Close: start your prompt library (10 min)

Each student starts a Notes / doc / wherever-you-keep-stuff file titled "My Prompt Library." They add their first three entries from this unit:

1. The strict-editor self-critique loop.
2. The chain-of-thought math template.
3. One persona of their own choosing.

Hand-wave the rest of the year: by the end of the course, this file should have 20–40 entries. The prompt library is the single most valuable artifact a student leaves this course with.

Differentiation, IEP, and 504 supports

• Executive function support: prompt chaining is, structurally, an executive function aid. Students with ADHD often find chains more useful than monolithic prompts.
• Read-aloud students: every prompt and persona can be voice-dictated.
• Advanced students: assign the meta-prompt - "design a system prompt that turns ChatGPT into a debate coach for me." Have them iterate it across two LLMs.

Assessment & evidence

• Formative: CoT vs. no-CoT scoring sheet, self-critique-loop reflection, chain artifact.
• Summative: quiz (10 questions). Three entries in the personal prompt library, screenshotted or pasted as a portfolio submission.

What's next

Unit 5 surfaces from technique to landscape: the big LLMs students hear about - GPT, Claude, Gemini, Llama, Mistral - what their actual differences are, what they cost, when to pick which.