Prompt Design — Intermediate

Student working on intermediate GenAI prompt design If you get stuck during this in-class exercise, ask the instructor. Let’s level up.

What you’ll learn

  • A quick CRAFT refresh and how to extend it with constraints/evidence.
  • Reusable prompt patterns (Summarizer, Explainer, Tutor, Planner, Critic, Formatter).
  • Prompt chaining to refine results and build on previous steps.
  • How to control tone, scope, and structure so outputs are audibly useful.
  • How to spot/avoid common failure modes (vagueness, missing evidence, scope creep).

Privacy note: Don’t paste confidential or personal data. For example, redact any personally identifying information for example:

  • Names
  • Emails
  • Street or postal code Addresses
  • Birthdates or other significant dates for events in people’s lives

  • GPS data in photos

CRAFT (fast refresh)

Use CRAFT and add constraints/evidence when facts matter.

  • Context — why you’re asking; Role — who to act as; Action — exact deliverable;
    Format — headings/bullets/JSON; Target Audience — who it’s for.
  • Constraints (word/time limits, must include/avoid) and Evidence (citations/URLs).

Typo fix from earlier pages: “Target Audience,” not “Target Audience.” Accuracy starts with you.

Plus:

  • Constraints: limits + must include/avoid (reduces errors + fluff)
  • Evidence: citations/URLs when facts matter

  • Acceptance Criteria (the “definition of done”): what a correct answer must include

Patterns library (copy/paste)

Use these as building blocks; swap in your topic.

1) Summarizer

Prompt

  • Context: You are summarizing for a reader who needs only what is supported by the text.
  • Role: Academic editor
  • Action: Summarize the text into 140–170 words.
  • Format: Headings: Key claim | Evidence used | Limitations | What is NOT stated
  • Constraints: No new facts. If missing, say “NOT PROVIDED.”
  • Evidence: Include 2 citations or quotes (<=10 words each).

2) Explainer

Prompt

  • Role: Subject expert
  • Action: Explain <concept> for <audience>.
  • Format: 5 bullets: Definition | Why it matters | Core idea | Simple example | Common mistake
  • Constraints: Plain language; avoid jargon; <=120 words.

3) Planner

Prompt

  • Role: Project planner
  • Action: Produce a 5-step plan to achieve <goal>.
  • Format: Table: Step | Objective | Inputs | Deliverable | Acceptance criteria
  • Constraints: No step longer than 2 lines.

4) Critic

Prompt

  • Role: Critical reviewer
  • Action: Assess my draft for Accuracy, Coverage, Clarity, Evidence, and Formatting.
  • Format: Table: Category | Score(1–5) | Evidence from text | Fix
  • Constraints: Do not rewrite yet. Only diagnose.

5) Formatter

Prompt

  • Role: Technical formatter
  • Action: Reformat this into Markdown with H2/H3 and numbered steps.
  • Constraints: Do not change meaning. Flag missing refs as [MISSING REF].

6) Template pattern (placeholders)

Prompt Use this exact template. Do not add extra sections: Title: Goal (1 sentence): Key points (3 bullets): Risks/uncertainties (2 bullets): Next actions (3 numbered steps): —

Prompting (the “3-step upgrade”)

Chaining beats one giant prompt. You get less chaos and more control.

Chain format: 1) Draft the output
2) Critique it using acceptance criteria
3) Revise based on critique

Example chain (copy/paste)
Step 1: Draft Create a structured answer on <topic>. Format: 6 bullets with headings. Include 2 sources.

Step 2: Critique `Check your answer against these acceptance criteria:

  • Has a clear definition (1 sentence)
  • Has 2 verifiable sources (URLs)
  • No claims without evidence
  • Max 120 words Return a table: Criterion | Pass/Fail | Fix`

Step 3: Revise Rewrite the answer applying all fixes. If you cannot verify a claim, mark it “NOT SURE.”

Advanced tips & examples

1) Start a new chat when context changes
Copilot’s New Topic button

  • Old context can contaminate answers. Use New Topic (or “new chat”) for unrelated tasks.

2) Tone control

  • Poor: Weather in Paris.
  • Better: Give me a lighthearted weather update for Paris, France, with a humorous twist.
  • Now try: same request in neutral and formal tones. Which suits your goal?

3) Examples & analogies (few-shot prompting)

  • Poor: Explain cyclones.
  • Better: Explain how cyclones form, using Cyclone Nisarga as the example. Provide sources.
  • Try adding one mini example of your own to guide the style.

4) Limit scope

  • Poor: Tell me everything about weather.
  • Better: Outline 4 factors that influence thunderstorm formation; 1 sentence each; list sources.
  • Scope: fewer errors, faster checking.

5) Iterate with acceptance criteria

  • Initial: Weather in London.
  • Refined: Weekend forecast for London, UK, with any weather warnings. Format: bullets; include source links.
  • Add acceptance criteria: “Contains date range, temperature range, precipitation chance; max 80 words.”

6) Prompt chaining (build on prior output)

  • First: What factors affect local weather conditions?
  • Then: Considering those factors, draft a 3-bullet safety advisory for hikers near San Francisco this weekend. Include sources.
  • Chaining = focused refinement, not rambling.

7) Leverage precedents (transfer learning) AI-generated student learning about weather

  • Prior prompt: Explain El Niño and its impact on global weather patterns.
  • New prompt: Explain La Niña and its impact on global weather patterns. Provide sources.

8) Assign a useful role

  • Role: You are the head of a creative department at an ad agency.
  • Action: Brainstorm 5 campaign taglines for <product>.
  • Constraints: Family-friendly; avoid superlatives; 6 words max per tagline.

9) Image generation (availability varies by tool/tier)

  • Try Meta.ai or Copilot.
  • Prompt: Create an image of a Grade 2 student learning about the weather in a classroom.
  • Constraints: Check for artifacts (extra fingers, text gibberish, anatomy errors) before using.

Badge evidence: capture a screenshot of your own image prompt + output.

Practice set (15–20 min)

A. Global warming (text)

  • Initial: What is global warming and what are its causes? Provide sources.
  • Scope: Limit to 300 words at a Grade 4 reading level.
  • Tone (style shift): Reword in a playful style suitable for Grade 4, while keeping facts accurate and cited.

(Avoid asking for an identifiable author’s proprietary style; keep it generic.)

B. Global warming (image)

  • Prompt: Create an image of a Grade 4 student learning about global warming in a classroom. AI-generated drawing of a student learning about global warming
  • Variation: Adjust the illustration to look like simple children’s book line art.

C. Lesson plan chaining

  • Start a new chat, then: I teach Grade 4. Generate a 60-minute lesson on local weather with a 30-minute hands-on activity.
  • Format: Objectives, Materials, Steps with times, Assessment, Differentiation.
  • Follow-up: Provide step-by-step student instructions for the activity at a Grade 4 reading level.
  • Add acceptance criteria: The plan must fit 60 minutes total, list 3 materials only, and include one formative check.

D. Pattern remix (make a reusable prompt) Pick a topic you actually care about (course, work, hobby). Copy/paste and fill blanks: **Prompt Context: I need a clean, checkable explanation for <audience>. Role: You are a <role>. Action: Explain <topic>. Format: 5 bullets: Definition | Why it matters | Core idea | Example | Common mistake. Constraints: <=120 words, plain language, no jargon. Evidence: Include 2 URLs if facts are involved; otherwise say “No sources required.”

E. Chain it (draft → critique → revise) Use the 3-step chain on ANY topic. **Step 1 Draft a structured answer about <topic>. Format: Headings + bullets. Include 2 sources (URLs). Step 2 Critique using this rubric (score 1–5 each): Accuracy | Completeness | Structure | Evidence | Audience-fit Return a table: Category | Score | What failed | Fix Step 3 Rewrite applying all fixes. Constraints: <=140 words. If unsure about a claim, mark it “NOT SURE.”

**F. Tone + scope A/B test (control the output) Run both prompts and compare.

A (neutral + tight scope) Explain <topic> in a neutral tone. Scope: only 3 key points, 1 sentence each. Format: bullets.

B (friendly + broader scope) Explain <topic> in a friendly tone. Scope: 6 key points, include 1 example. Format: bullets.

Reflection (2–3 min)

  • Which acceptance criteria improved quality the most?
  • Where did tone or scope reduce errors?
  • What will you reuse as a personal template?

Self-check (2 min)

  • Did you specify CRAFT + constraints/evidence?
  • Is tone and scope appropriate for your audience?
  • Did you chain prompts rather than rewrite from scratch?
  • Are sources verifiable (and not just names without links)?
  • Would your output pass your own acceptance criteria?

Acceptance criteria mini-library

  • Summary: 120–150 words, neutral tone, no new facts, 1 quote ≤10 words.
  • Plan: 5 steps; each has Objective • Inputs • Acceptance; total ≤180 words.
  • Table: Columns fixed; max 6 rows; no empty cells.
  • Evidence: Each claim has URL + date; if unsure → “NOT SURE”.

Go further

A practical overview of techniques worth skimming next:
Prompt engineering techniques

NEXT STEP: Idea Generation