Homework

What to do

Build a 2-step prompt chain. Step 1 extracts facts from a provided 500-word article. Step 2 answers questions using only those extracted facts — not from general knowledge.

Requirements

• Find or write a 500-word factual article (technology, science, or business)
• Step 1 prompt: extract all key facts as a numbered list
• Step 2 prompt: answer 5 questions using only the extracted facts; if a fact is not present, say "Not in the article"
• Write 5 questions — at least 2 should be answerable from the article and at least 1 should not be
• Verify that Step 2 correctly says "Not in the article" for the unanswerable question

facts = call_claude(f"Extract all key facts from this article as a numbered list:\n\n{article}")
answers = call_claude(
    f"Answer each question using ONLY the facts listed below. "
    f"If the answer is not in the facts, say 'Not in the article'.\n\n"
    f"FACTS:\n{facts}\n\nQUESTIONS:\n{questions}"
)

Self-check: Does the Step 2 answer for the unanswerable question correctly say "Not in the article"? If not, your Step 2 prompt may be allowing general knowledge to leak in.

What to do

Build a few-shot classifier for a domain relevant to your work or interests. The classifier must have at least 5 categories and use at least 3 labeled examples per category in the prompt.

Requirements

• At least 5 categories (e.g., bug severity levels, email types, article genres, support ticket types)
• At least 3 examples per category in the few-shot prompt (15+ examples total)
• Test on 10 samples — at least 5 must be correctly classified
• Document: category definitions, example selection rationale, accuracy result

Example domains

• Code review comments → categories: architecture/bug/style/performance/security
• News articles → categories: politics/technology/sports/business/science/culture
• Customer feedback → categories: bug/feature request/praise/complaint/question

Self-check: Did you include diverse examples (not all similar)? Does your classifier output exactly the category label with no extra text?

What to do

Design a complete system prompt for a customer service chatbot for a product of your choice. The system prompt must define a persona, 5 behavioral constraints, and a response format. Then test whether the bot refuses off-topic requests.

Requirements

• Persona: Name, role, company, tone (1–2 sentences)
• 5 behavioral constraints: Things the bot must always or never do. Examples:
  • "Never reveal pricing before the user has described their use case."
  • "Always offer to connect the user to a human agent if frustrated."
  • "Never discuss competitors by name."
• Response format: Max length, structure, whether to use lists or prose
• 5 test messages: 3 on-topic (should be answered), 2 off-topic (should be declined politely)

Scoring your system prompt

Aim for 5/5. Document any failures and what you changed to fix them.

Self-check: Did the bot stay in character across all 5 tests? Did it use the correct tone? Did the off-topic refusals avoid being rude?