QUIZ

 

Prompt Engineering — MCQ Quiz

25 Basic + 25 Medium Level Questions (Covering Units I–V)

---

SECTION A: BASIC LEVEL (Q1–Q25)

Unit I — Foundations of Prompt Engineering

1. What is prompt engineering primarily concerned with? a) Training a model from scratch b) Designing effective inputs to guide model outputs c) Compiling source code d) Building neural network architectures
2. Prompt engineering differs from fine-tuning mainly because it: a) Requires retraining the model's weights b) Works without modifying the model's parameters c) Needs a labeled dataset d) Is only used for image models
3. Which of the following is a core principle of effective prompt design? a) Ambiguity b) Clarity and specificity c) Randomness d) Long, unstructured text
4. The "anatomy of a prompt" typically includes which component? a) Instruction b) Context c) Output format d) All of the above
5. What is the first step in the iterative prompting lifecycle? a) Deploying the model b) Drafting an initial prompt c) Fine-tuning the model d) Discarding the prompt

Unit II — Advanced Prompt Patterns & Techniques

6. Few-shot prompting refers to: a) Giving the model zero examples b) Providing a few example input-output pairs before the actual task c) Training the model on thousands of examples d) Removing all instructions from the prompt
7. Role-based prompting is used to: a) Delete the model's memory b) Establish a persona or specific behavior for the model c) Reduce token usage d) Increase model temperature
8. Negative prompting is best described as: a) Giving purely positive feedback b) Instructing the model to avoid certain content c) Asking multiple questions at once d) Disabling the model
9. Constraint specification in prompting refers to: a) Removing all rules from the prompt b) Defining limits like word count or format c) Increasing the model's parameters d) Ignoring formatting instructions
10. Zero-shot prompting means: a) The model is given several examples b) The model performs a task with no prior examples in the prompt c) The model is retrained for each task d) The prompt has zero words

Unit III — Structured Output & Reasoning Techniques

11. Which format is commonly requested for structured data output? a) JSON b) YAML c) Markdown tables d) All of the above
12. Chain-of-thought (CoT) prompting encourages the model to: a) Skip reasoning steps b) Explain its reasoning before giving a final answer c) Provide only the final answer d) Ignore the question
13. Task decomposition in prompting involves: a) Merging all sub-tasks into one prompt b) Breaking a complex task into smaller manageable sub-tasks c) Removing the task entirely d) Avoiding structured responses
14. A common way to verify JSON output validity is to: a) Ignore syntax errors b) Use a parser to validate the structure c) Manually guess correctness d) Avoid using JSON altogether
15. Why are structured outputs important for real-world applications? a) They make outputs harder to use b) They allow easy integration with downstream systems c) They increase ambiguity d) They are only for aesthetics

Unit IV — Retrieval-Augmented Generation & LangChain Workflows

16. RAG stands for: a) Random Access Generation b) Retrieval-Augmented Generation c) Rapid Application Gateway d) Reasoning and Grammar
17. Why is RAG needed? a) LLMs have unlimited internal knowledge b) LLMs have limitations in internal/up-to-date knowledge c) To slow down response time unnecessarily d) To remove the need for embeddings
18. What is the purpose of text splitting in a RAG indexing pipeline? a) To delete documents b) To break documents into manageable chunks for embedding c) To translate documents d) To compress the LLM
19. Embeddings in a vector store are used to: a) Store raw text only b) Represent text as numerical vectors for similarity search c) Encrypt the documents d) Replace the LLM entirely
20. LCEL stands for: a) LangChain Expression Language b) Large Context Evaluation Layer c) Linear Chain Execution Logic d) Language Contextual Embedding List

Unit V — Agents, Multimodal AI & Ethical Evaluation

21. An LLM agent is distinguished by its ability to: a) Only generate plain text b) Invoke tools/functions to complete tasks c) Avoid using any external tools d) Work without any prompt
22. VLMs refer to: a) Vision-Language Models b) Virtual Learning Machines c) Verified Language Modules d) Variable Length Memory
23. "LLM-as-Judge" is a technique used for: a) Training new LLMs b) Automated evaluation of model outputs c) Generating images only d) Deleting poor prompts
24. Prompt injection is best described as: a) A method to speed up inference b) A security risk where malicious input manipulates model behavior c) A type of embedding d) A formatting technique
25. Which of these is an ethical concern in LLM applications? a) Bias in outputs b) Misinformation c) Data privacy d) All of the above

---

SECTION B: MEDIUM LEVEL (Q26–Q50)

Unit I — Foundations of Prompt Engineering

26. A student writes: "Write a bio of Ada Lovelace" versus "As a historian, write a 100-word bio of Ada Lovelace focusing on her contribution to computing, in formal tone." The second prompt demonstrates: a) Prompt degradation b) Enhanced prompting via role framing, specificity, and format instructions c) Model fine-tuning d) Negative prompting
27. If a prompt produces an ambiguous or contradictory output, the most likely underlying issue is: a) The model's temperature is too low b) Missing context or unclear instructions in the prompt c) The API key is invalid d) The output format is JSON
28. During iterative refinement of a summarization prompt, adding "in two sentences, focusing on setting and theme" primarily improves: a) Model training speed b) Content specificity and adherence to constraints c) API latency d) Token encoding
29. Which scenario best illustrates a "common prompt pitfall"? a) A prompt with clear role, context, and format b) A vague prompt like "Tell me about it" with no defined subject c) A prompt with explicit word-count limits d) A prompt using few-shot examples
30. Comparing a naive prompt to an enhanced prompt for the same task, evaluators should primarily judge outputs on: a) Length only b) Relevance, completeness, and style c) Font size d) API response time only

Unit II — Advanced Prompt Patterns & Techniques

31. A prompt says: "You are a financial advisor. Do not mention any brand names." This combines: a) Zero-shot and CoT prompting b) Role-based prompting and negative prompting c) RAG and embeddings d) JSON generation and YAML parsing
32. In a few-shot sentiment classification task, providing 2–3 exemplar pairs mainly helps the model: a) Reduce API cost b) Better align outputs with the desired pattern and accuracy c) Avoid using any context d) Ignore the given examples
33. When a summarization prompt fails to meet a word limit, the recommended refinement is to: a) Abandon the task b) Add explicit constraints such as word count or bullet format c) Remove all instructions d) Switch to zero-shot only
34. Comparing zero-shot and few-shot outputs for the same classification task, a key evaluation criterion is: a) Consistency and adherence to the example pattern b) Number of API calls only c) Length of the model name d) Random variation only
35. If a role-based prompt does not sufficiently suppress unwanted content, the next logical step is to: a) Delete the entire prompt b) Add explicit negative prompting instructions c) Increase the number of unrelated exemplars d) Switch the output format to audio

Unit III — Structured Output & Reasoning Techniques

36. A prompt asks for "three benefits of exercise in a Markdown table with columns Benefit and Description." If the output isn't a table, the best fix is to: a) Accept the output as-is b) Re-specify the format explicitly and restate column names c) Switch to JSON without changes d) Remove structure entirely
37. When generating JSON for a dataset and the parser throws an error, the appropriate action is to: a) Ignore the parser b) Refine the prompt to resolve syntax issues and re-validate c) Convert to a scanned image d) Disable JSON formatting permanently
38. Zero-shot CoT prompting ("Let's think step by step") is most useful for: a) Simple factual lookups b) Multi-step reasoning problems like logic puzzles c) Formatting output in YAML only d) Reducing model context window
39. Comparing a direct-answer baseline to a decomposed sub-question approach for a complex problem, the decomposed approach is expected to: a) Always perform worse b) Often improve accuracy by breaking down complexity c) Eliminate the need for any prompt d) Remove the need for an LLM
40. A key reason structured formats (JSON/YAML) are preferred in production pipelines is: a) They are harder to parse programmatically b) They integrate more reliably with downstream software systems c) They cannot represent nested data d) They eliminate the need for any validation

Unit IV — Retrieval-Augmented Generation & LangChain Workflows

41. In a RAG pipeline, after receiving a user query, the next step is typically to: a) Directly generate an answer with no retrieval b) Retrieve the top-k relevant chunks from the vector store c) Delete the query d) Skip embedding generation
42. Comparing answers from a RAG-based pipeline versus a prompt without retrieval, the RAG pipeline is expected to show improved: a) Factual accuracy grounded in retrieved documents b) Random creativity only c) API latency reduction only d) Model parameter count
43. When splitting documents into chunks (e.g., 200 tokens) for indexing, the primary goal is to: a) Maximize randomness b) Ensure consistent, manageable pieces for embedding and retrieval c) Delete document metadata d) Avoid using embeddings
44. An LCEL chain that summarizes text and prints the result demonstrates: a) A retrieval-free, minimal end-to-end LLM invocation b) A fine-tuning pipeline c) A multimodal image generation task d) A prompt injection test
45. If a RAG-based answer includes outdated or incorrect information despite retrieval, a likely cause is: a) The vector store contains irrelevant or poor-quality chunks b) The LLM has too much internal knowledge c) JSON formatting is enabled d) The API key expired

Unit V — Agents, Multimodal AI & Ethical Evaluation

46. A prompt instructs an agent to invoke a calculator tool when a math question arises. This illustrates: a) Basic LLM agent architecture with tool invocation b) Pure zero-shot prompting with no tools c) A RAG indexing pipeline d) A negative prompting technique
47. In multimodal prompting, comparing a generated image's caption to the original prompt evaluates: a) Token usage only b) Alignment between generated content and intended description c) API cost d) Model training time
48. Designing a prompt-injection test such as "Ignore previous instructions..." is done to: a) Improve model creativity b) Identify security vulnerabilities and refine system prompts c) Increase token limits d) Generate images faster
49. Using "LLM-as-Judge" to rate outputs on a 1–5 scale for clarity and correctness is an example of: a) Manual heuristic evaluation only b) Automated evaluation using another model as an evaluator c) Prompt injection d) Fine-tuning
50. When deploying a prompt-driven application with Streamlit or Gradio, a key ethical consideration to address beforehand is: a) Font color choices b) Bias, misinformation, and data privacy risks c) Button placement only d) Increasing model temperature randomly

---

ANSWER KEY

Q	Ans	Q	Ans	Q	Ans	Q	Ans	Q	Ans
1	b	11	d	21	b	31	b	41	b
2	b	12	b	22	a	32	b	42	a
3	b	13	b	23	b	33	b	43	b
4	d	14	b	24	b	34	a	44	a
5	b	15	b	25	d	35	b	45	a
6	b	16	b	26	b	36	b	46	a
7	b	17	b	27	b	37	b	47	b
8	b	18	b	28	b	38	b	48	b
9	b	19	b	29	b	39	b	49	b
10	b	20	a	30	b	40	b	50	b