Prompt Engineering Interview Questions: Zero-Shot, Few-Shot, CoT, ReAct & More

What is Prompt Engineering?

• Directly used by end users.
• If LLM understands human language, then why do we need to learn how to interact with LLM?
  • Because best input gives best output.
• Prompt Engineering is a way of guiding the LLM to respond properly.
  • It is the art and science of crafting effective inputs (called prompts) to guide an LLM or other GenAI model to produce desired, accurate, and relevant output.
  • It’s about optimizing communication between human and AI.

Types of Prompt Engineering Techniques

1. Zero-shot prompting

2. One-shot prompting

3. Few-shot prompting

4. Role-play prompting

5. Chain of Thought (CoT) prompting

6. ReAct prompting (Advanced technique)

7. Meta prompting (Advanced technique)

   (and constantly evolving...)

1. Zero-Shot Prompting

• Direct question to a model without any example or format.
• Used when you expect a direct answer from the LLM.
• Example: "Who is Modi ji?"

2. One-Shot Prompting

• Where you provide both the question and answer, showing a format or pattern.
• Example:
  • Q: What is 2 + 2?
  • A: The answer is 4.
• So next time it will answer in the same way:
  • Q: 4 + 3
  • A: The answer is 7 (followed the format).

3. Few-Shot Prompting

• Extension of one-shot prompting where you provide more than 1 example.
• Used when various cases need to be considered.
• Example:
  • "The movie was good" → Positive
  • "I was sad" → Negative
  • "Weather?" → Neutral

4. Role-Play Prompting

• Introducing the model to adopt a specific persona or role.
• Useful when the task is complex and a specific role can help.
  • Want to know about a dish → Chef persona
  • Need customer support → Customer service agent
  • Academic explanation → Professor
• Example:
  • "You are a historian specialized in WWII. Explain the significance of Battle X."
  • You could ask this directly, but adding this persona gives a much more tailored answer.

5. Chain of Thought (CoT) Prompting

• Encouraging the model to explain its reasoning process step by step before arriving at the final answer.
• Used for complex reasoning tasks, mathematical problems, and multi-step queries.
• Format:
  • Q: A descriptive/reasoning question
  • A: Step-by-step reasoning → Final answer

6. ReAct Prompting (Reason + Act)

• Method where the LLM performs more complex tasks by adding explicit steps through interaction with external tools and the environment.

• The LLM goes through a repeated cycle until the final result is reached:

  Thinking → Action → Observation

• Example:

  • Q: What is the capital of Japan and its population?
  • Thought: Need to find out the capital and population.
  • Action: Search (capital of Japan)
  • Observation: Tokyo
  • (Cycle repeats for the second question)
  • Action: Search (population of Tokyo)
  • Final Answer: The capital of Japan is Tokyo and the population is around 14 million.

7. Meta Prompting

• Advanced prompt engineering technique where you essentially use an LLM to generate, refine, or optimize other prompts.
• When to use:
  • When you have a short/rough prompt and ask the LLM to act as a prompt engineer to expand, add detail, and make it a comprehensive, high-quality prompt.

8. Temperature Parameter

• A parameter that controls the randomness and creativity of the model’s output.
• Normally scaled from 0 to 1 (or up to 2).
• Low Temperature (Closer to 0):
  • Deterministic, conservative, predictable, and consistent.
  • Use cases: Summarization, code generation, translation.
• High Temperature (Closer to 1 or above):
  • Diverse, creative, random.
  • Use cases: Poetry, brainstorming, creative writing.

9. Top-P and Top-K Sampling

Parameters that control which tokens are considered when the model generates output, directly affecting creativity and randomness (fine-tuning beyond temperature).

• Example sentence: "The cat sat on the ______"
  • Low Temperature: High-probability tokens like mat (0.53), rug (0.40).
  • High Temperature: Lower-probability words like pizza, aircraft, spacecraft also get considered.

Top-K:

• Limits the sampling pool to strictly the top K most probable tokens.
• Example:
  • Word probabilities for "The color of cloud is ______":
    • White (0.29)
    • Grey (0.23)
    • Yellow (0.11)
    • Orange (0.09)
    • Red (0.08)
    • Black (0.06)
  • If Top-K = 5, it will only consider the top 5 values (White, Grey, Yellow, Orange, Red) and discard the rest.

Top-P (Nucleus Sampling):

• Selects from the smallest group of tokens whose cumulative probability exceeds the threshold P.
• Example:
  • If Top-P = 0.90, it adds probabilities starting from the highest until the cumulative sum hits 90% (0.90), cutting off the long tail of unlikely words.