Unit 3: Prompt Engineering Fundamentals
Lesson at a glance
| Item | Detail | | --------------------- | ------------------------------------------------------------------ | | Suggested length | 3 × 60 minutes | | Recommended placement | Week 3 of AI Fluency | | Prerequisite | Units 1–2 complete | | Materials | Student devices with frontier-LLM access, prompt-evaluation rubric |
Safety: Students will run prompts. Reinforce: never paste private data, never paste another student's work, never paste copyrighted material wholesale.
Standards & credential alignment
- AI4K12 Big Ideas: Natural Interaction, Representation & Reasoning.
- CSTA K-12: 3A-CS-01, 3A-AP-13, 3A-IC-25.
- ISTE Students: 1.5 Computational Thinker, 1.6 Creative Communicator.
Learning objectives
By the end of this unit, students can:
- Apply the C.R.I.S.P. frame (Context, Role, Instructions, Specifics, Polish) to any prompt.
- Distinguish system prompt vs. user prompt and use each appropriately.
- Use zero-shot, one-shot, and few-shot prompting and explain when each is best.
- Force structured output (JSON, table, bullet list) reliably.
- Iterate a weak prompt into a strong one in 3–5 rounds using a published rubric.
- Recognize and avoid five common prompt anti-patterns.
Vocabulary
- System prompt - Instructions given to the model before the conversation begins. Sets persona, rules, format. Hidden from the user in most chat apps.
- User prompt - What you type. The thing you have direct control over.
- Zero-shot - Prompt with no examples. ("Translate this to French: …")
- Few-shot - Prompt with 2–5 examples of the pattern you want. Dramatically improves consistency.
- Chain-of-thought - Telling the model to "think step by step" before answering. (Deeper coverage in Unit 4.)
- Persona - The voice and stance the model adopts ("You are a high school history teacher…").
- Constraint - A hard limit you set ("Respond in 50 words or fewer." "Cite only sources I provide.").
- Format spec - How you want the answer structured. JSON, markdown table, bulleted list, etc.
- Anti-pattern - A prompt habit that consistently produces worse results (the unit's "what not to do").
The C.R.I.S.P. frame
Drop this on the wall and leave it for the rest of the year:
Context - Who is asking, who's it for, what's the situation. Role - Who should the AI be? ("You are a patient calculus tutor.") Instructions - What you want done, in imperative voice. Specifics - Examples, constraints, must-haves and must-not-haves. Polish - Format, tone, length.
Compare two prompts on the same task:
❌ Weak prompt: "Help me with my essay on climate change."
✅ C.R.I.S.P. prompt:
Context: I'm a 10th-grade student writing a 5-paragraph argumentative essay on climate change for my AP Environmental Science class. The essay is due Friday and my draft is below.
Role: You are a strict but kind English teacher who reviews student writing for argument quality, not grammar.
Instructions: Read my draft. Tell me the three weakest spots in my argument and one specific way to strengthen each.
Specifics: Do not rewrite my essay. Do not give me new sources I haven't cited. Reference the paragraph number when you point to weaknesses.
Polish: Respond as a numbered list. Each weakness ≤ 30 words. Each suggestion ≤ 30 words. End with one sentence of encouragement.
Draft: [student pastes draft]
The second prompt is twice as long and produces ten times better feedback. That is the entire unit.
Pacing - Day 1 (60 min): C.R.I.S.P. and one-shot
| Time | Segment | Notes | | ----------- | ------------------------ | ---------------------------------------------------------------------- | | 0:00 – 0:05 | Recap | Tokens, attention, hallucination. Cold-call. | | 0:05 – 0:25 | Mini-lesson - C.R.I.S.P. | Compare weak vs. strong prompt live. | | 0:25 – 0:50 | Activity - Prompt Battle | Pairs run weak and C.R.I.S.P. version of same task. Score with rubric. | | 0:50 – 1:00 | Discussion | What changed when you added each letter? |
Day 1 - Activity: Prompt Battle (25 min)
Worksheet provides 4 task scenarios:
- Get feedback on a paragraph of writing.
- Generate flashcards for studying a chapter.
- Plan a 5-day study schedule for finals.
- Brainstorm science fair ideas.
Each pair runs the weak version (one sentence, no structure) and the C.R.I.S.P. version on the same task. They score both with the rubric (clarity, completeness, usefulness, format) and report the delta. The class average delta is usually 2–3x. That number lands the lesson harder than any explanation.
Pacing - Day 2 (60 min): Few-shot, format control, system prompts
| Time | Segment | Notes | | ----------- | ----------------------------------- | ----------------------------------------------------------------- | | 0:00 – 0:15 | Mini-lesson - zero / one / few-shot | When examples beat instructions. | | 0:15 – 0:30 | Mini-lesson - forcing structure | JSON / table / list output. | | 0:30 – 0:50 | Activity - Few-shot extraction | Pairs build a few-shot prompt that extracts data from messy text. | | 0:50 – 1:00 | Discussion - when to add an example | |
Day 2 - Mini-lesson: zero / one / few-shot (15 min)
The rule of thumb: if a task is unusual or has a specific style, give examples. If it's a common task, just describe it.
Few-shot prompt example for sentiment classification:
Classify the sentiment of each sentence as POSITIVE, NEGATIVE, or NEUTRAL.
Example 1:
Sentence: "The pizza was cold and the service was slow."
Sentiment: NEGATIVE
Example 2:
Sentence: "I had the salad. It came on a plate."
Sentiment: NEUTRAL
Example 3:
Sentence: "This is the best taco I have ever eaten in my entire life."
Sentiment: POSITIVE
Now classify:
Sentence: "[student input]"
Sentiment:
A model that struggles to classify sentiment zero-shot will nail it three-shot. The lesson: examples are how you teach the model new tricks without retraining it.
Day 2 - Mini-lesson: forcing structure (15 min)
Three reliable techniques:
- Show the format. "Reply in this exact JSON shape:
{"name": "", "score": 0, "notes": ""}" - Use a delimiter. "Wrap your final answer in
<answer></answer>tags." - Tell it to be a parser. "You are a function that takes English and returns a single JSON object. No prose. No code fences. Just the JSON."
Demo all three. Show what happens when you don't (markdown code fences sneak in, prose preamble, etc.).
Day 2 - Activity: Few-shot extraction (20 min)
Students get a paragraph of messy school-event text:
"The robotics meeting next Tuesday at 3:30 in room 204 has been moved to Wednesday at 4 in the cafeteria - coach Allen is out. Drama club is still Thursday at 3 in the auditorium. Saturday's game is canceled."
They build a few-shot prompt that returns:
[
{
"event": "robotics",
"day": "Wednesday",
"time": "4:00 PM",
"location": "cafeteria",
"status": "rescheduled"
},
{
"event": "drama club",
"day": "Thursday",
"time": "3:00 PM",
"location": "auditorium",
"status": "scheduled"
},
{
"event": "Saturday game",
"day": "Saturday",
"time": null,
"location": null,
"status": "canceled"
}
]
This is the same skill that powers most "AI does my paperwork" use cases.
Pacing - Day 3 (60 min): Iteration, anti-patterns, capstone
| Time | Segment | Notes | | ----------- | -------------------------------- | ----------------------------------------------------------- | | 0:00 – 0:15 | Mini-lesson - the iteration loop | Critique your own output. Ask the model to critique itself. | | 0:15 – 0:30 | Mini-lesson - five anti-patterns | What not to do. | | 0:30 – 0:55 | Activity - Prompt Olympics | Three rounds, scored. Best prompt wins. | | 0:55 – 1:00 | Quiz / exit ticket | |
Day 3 - Mini-lesson: the iteration loop (15 min)
Most students treat the first response as the answer. Pros treat it as a draft.
The loop:
- Run the prompt.
- Read the answer. Identify what's wrong (too long, wrong tone, missed a constraint, hallucinated, etc.).
- Don't rewrite from scratch. Add a follow-up: "Good - now make it half as long and remove the bullet about X."
- Repeat 2–4 times.
The model can critique itself. Add: "Before responding, list three ways your previous answer could be improved. Then write the improved version." This single trick measurably improves output quality on most reasoning tasks.
Day 3 - Mini-lesson: five anti-patterns (15 min)
| # | Anti-pattern | Why it fails | | --- | ------------------------------------------- | --------------------------------------------------------------------------------- | | 1 | "Be creative." | Vacuous. Tell the model which dimensions to vary. | | 2 | "Don't make mistakes." | Useless. The model doesn't know it's making one. Add examples of correctness. | | 3 | "Tell me everything you know about X." | Triggers a fluffy info-dump. Ask a specific question. | | 4 | Stacking 11 contradictory rules. | The model picks 3 and silently drops the rest. Cut to the 3 that matter. | | 5 | Pasting a 10K-word doc with no instruction. | The model summarizes by default. Tell it what to do with the doc. |
Day 3 - Activity: Prompt Olympics (25 min)
Three head-to-head rounds. Each round, the teacher reads a task. Students have 3 minutes to write the best C.R.I.S.P. prompt they can. Run all submissions on the model. Class votes on which output is best. Winner announces their prompt aloud. This activity is genuinely competitive and students remember it for the rest of the year.
Differentiation, IEP, and 504 supports
- Read-aloud students: prompts can be voice-dictated; emphasize speaking the C.R.I.S.P. frame aloud as a draft technique.
- EL students: C.R.I.S.P. works in any language. Encourage prompts in the student's strongest language; the model translates internally.
- Students who are already power users: assign the "system prompt" extension - they design a custom GPT / Claude Project for a specific use case (study helper, code reviewer, debate partner) and document the system prompt.
Assessment & evidence
- Formative: prompt-battle deltas, few-shot extraction, Olympics rankings.
- Summative: end-of-unit quiz (12 questions). Capstone artifact: a polished prompt for one of their other classes, submitted with the iteration trail.
What's next
Unit 4 goes deeper: chain-of-thought prompting, role-play and persona, ReAct, self-critique, and the prompt-library habit professional users build over months.