AI Pathfinder (Ages 15-18)

Module 2 of 10

Module 02: Prompt Craft

4 min read777 words
What you'll learn
Structure prompts with a repeatable framework (CRISP).Compare zero-shot, one-shot, and few-shot—and know when examples win.Iterate one variable at a time and save winners.Draft a system prompt for a real use case (study coach, club, hobby).Maintain a tiny prompt library with dates and “what failed” notes.Respect school rules—scaffolding, not smuggling graded work.

"Clear thinking becomes clear writing." — William Zinsser. Muddy instructions → muddy answers. Prompting is learnable—like free throws, not magic.

Relatable: It’s 11 p.m., your first prompt is “help,” and the bot writes a novel. There’s a better way.

Time: ~45–75 minutes first pass + 1–2 hours this week for a small prompt pack.

Learning Objectives

By the end of this module, you will be able to:

  • Structure prompts with a repeatable framework (CRISP).
  • Compare zero-shot, one-shot, and few-shot—and know when examples win.
  • Iterate one variable at a time and save winners.
  • Draft a system prompt for a real use case (study coach, club, hobby).
  • Maintain a tiny prompt library with dates and “what failed” notes.
  • Respect school rules—scaffolding, not smuggling graded work.

1. Why Prompts Are a Real Skill

Models continue text from what you gave them. Vague in → average out.

Analogy: Same phone camera—random snaps vs adjusting light and framing.

Fun Fact: Tiny wording shifts change probabilities—not because the model has “favorite words,” but because you moved the next-token landscape.

2. CRISP Framework

Context · Role · Instruction · Style · Parameters (length, must-nots, safety).

In your notes, write a template with blanks for each letter—no need to memorize ours verbatim. Example shape: “10th grader, bio tutor, explain Calvin cycle in six steps with check questions, friendly tone, max 250 words, hints not full answers first.”

Try This! Run the same task twice: (1) one vague sentence (2) CRISP. Paste both outputs side by side—label what improved.

3. Zero-Shot vs Few-Shot

TechniqueWhen
Zero-shotQuick, clear format
One-shotStyle matching
Few-shot“Questions like these → answers like these”

Concept: Showing a format beats describing it with extra adjectives.

Activity: Build three tiny examples yourself for something repetitive (vocab, club posts). Teaching the pattern to the model teaches you what you want.

4. System Prompts = House Rules

Often include: role, audience, integrity line (“don’t complete graded submissions”), output shape, uncertainty habit (“say if unsure”).

Try This! Write two system prompts: strict coach vs gentle mentor. Same user message—notice tone shift.

5. Chain-of-Thought (Carefully)

“Asking for steps” helps you inspect reasoning—not guarantee correctness. Circle the step you’d ask a teacher about.

6. Iteration Loop

  1. Run → 2. Edit-like read → 3. Fix one knob → 4. Re-run → 5. Save winner.
SymptomFirst fix
Too longMax words / bullet cap
Too genericAudience + stakes + example
Wrong toneStyle line + one-shot
Makes stuff upAsk for uncertainty + check plan
Ignores formatRepeat format at end

Try This! Same task, three prompts—each improves one measurable thing (length, structure, tone). Keep the evolution in a doc.

7. Prompt Library (Simple)

Fields: name, full prompt, tool tested, what worked, what failed, date (models update).

Fun Fact: Versioning (v1, v2) sounds nerdy until it saves you on deadline night.

8. Settings When You See Them

Temperature: higher → more varied; lower → steadier. Max length: always set for school tasks so you don’t drown.

Try This! Same question: 80-word cap vs no cap—which helped you learn more?

9. When Not to Prompt

Personal essays need your specifics; legal/medical stakes need pros; forbidden exam use; mystery citations for graded claims.

Practice Challenges

Five CRISP prompts for real life; tone matrix on one factual topic; few-shot note converter; system prompt duel with a friend; failure autopsy on an old bad output; integrity block for custom instructions.

Your Challenge

Prompt Pack (10 entries): 4 school helpers (tutor, not ghostwriter), 2 creative, 2 productivity, 2 “fix my prompt” meta-prompts. Each: name, CRISP text, 2–5 line output snippet, when not to use.

Plus one-page “How I Iterate” with three-step checklist and one v1→v3 example.

Discuss with a friend: Can they use your pack without DMing you?

Key Takeaways

  • Prompting = instruction design; clarity beats cleverness.
  • Examples often beat adjectives.
  • System prompts set stable rules; chats handle the moment.
  • Iterate and date your library.

Key Takeaway

  • Use CRISP so you’re never staring at a blank box.
  • Change one variable per iteration and keep the winner.
  • Few-shot patterns reduce format fights fast.
  • System rules should include integrity boundaries for school.
  • For graded work, your understanding is the product—AI is scaffolding when allowed.

Going Further

Official tool docs; pair with Module 03 for fact-checking; share skeletons, not private data.

You’re learning a language machines respond to—with reps, you’ll sound like you meant it.