Syed Jafer K

Prompting is the process of giving specific instructions, questions, or examples to a Large Language Model (LLM) like ChatGPT to guide its output toward the desired goal.

In other words, its about how you are making conversation with an LLM.

Goal of Prompting

• To clarify the task for the AI.
• To control the format, tone, and depth of the output.
• To encourage reasoning or multi-step thinking.
• To reduce hallucination by narrowing the context.

Type	Description	Example
Zero-Shot	No examples given	“Summarize this text.”
One-Shot	One example given	“Example: Dog → perro; Cat → ?”
Few-Shot	Multiple examples	“Translate the following… cat→chat, dog→chien…”
Chain-of-Thought	Ask to reason step by step	“Think step by step.”
Role-Based	Assign an identity	“You are a Python teacher.”
Self-Reflexive	Ask model to review itself	“Check and improve your answer.”

Below are some examples , just for reference. I hope we are not moving to do a GridSearch to find which prompt works better ? Who Knows ???

1. Zero Shot Prompting

Ask the model to perform a task without prior examples.

Examples

1. “Write a poem about deep learning.”
2. “Summarize this paragraph in 2 lines.”
3. “Explain quantum computing in simple terms.”
4. “Generate a SQL query to find all customers who made purchases in the last month.”

2. One-Shot Prompting

Provide one example to show the expected format.

Examples

1. Example
   Q: What is the capital of Japan?
   A: Tokyo
   Q: What is the capital of Italy? ➜ Output: Rome
   
2. Example
   English → Spanish
   cat → gato
   dog → ? ➜ Output: perro
   
3. Example
   Input: “Increase brightness”
   Output: “Adjusting display to +10% brightness”
   Input: “Lower volume” → ? ➜ Output: “Decreasing sound by 10%”

3. Few-Shot Prompting

Give multiple examples to teach the model a consistent pattern.

Examples

1. Sentiment Classification
   • “I love this movie!” → Positive“This is terrible.” → Negative“It’s okay, not great.” → Neutral“The acting was amazing!” → ?
   ➜ Positive
   
2. SQL Generation
   • “List all employees.” → SELECT * FROM employees;“ List all products.” → SELECT * FROM products;“List all customers.” → ?
   ➜ SELECT * FROM customers;

4. Chain-of-Thought Prompting (CoT)

Ask the model to think step by step before answering.

Examples

1. “If 4 pens cost ₹20, how much do 10 pens cost? Think step by step.”
   ➜ Step: 1 pen = ₹5 → 10 × 5 = ₹50
   
2. “Why does the sun appear red during sunset? Explain step by step.”
   ➜ Step 1: Sunlight passes through thicker atmosphere → Step 2: Blue light scatters → Step 3: Red light reaches eyes.
   
3. “You have 12 marbles, give 4 to a friend. How many left? Explain.”
   ➜ Step: 12 − 4 = 8 marbles.

5. ReAct Prompting (Reason + Act)

Combine reasoning and action explicitly.

Examples

1. “You’re a travel planner. Think about ideal locations for a short vacation from Chennai. Then list 3 options.”
2. “You are an SRE. Think about why the API latency is high. Then suggest 3 mitigation steps.”
3. “You are a data analyst. Think about how to visualize daily user logins. Then produce the Matplotlib code.”

6. Self-Consistency Prompting

Ask for multiple reasoning paths to get the most reliable answer.

Examples

1. “What is 25% of 160? Give 3 reasoning paths and pick the best.”
   ➜ All lead to 40.
2. “How many minutes in 3.5 hours? Solve in 2 ways and choose consistent.”
   ➜ 3.5×60 = 210.
3. “If the train leaves at 3 PM and travels 2.5 hours, what time will it reach?”
   ➜ 5:30 PM via all methods.

7. Socratic Prompting

Use a question-driven approach to guide understanding.

Examples:

1. “What happens when we drop an object? Why? What is that force called?”
2. “If databases use indexes, what is their purpose? How do they help?”
3. “If a model overfits, what could that mean about its training data?”