Advanced Prompting: Zero-Shot to Chain-of-Thought

A Large Language Model is like a highly talented intern who has read the entire internet but has no idea what you specifically want from them today. To guide them, we use three primary "In-Context Learning" techniques.

In this lesson, we will explore Zero-Shot, Few-Shot, and Chain-of-Thought (CoT), and learn exactly when to use each in a production pipeline.

---

1. Zero-Shot Prompting (The Simplest Path)

Zero-Shot means you give the model a task with no examples. You are relying entirely on the model's pre-existing knowledge.

Use Case:

• Simple classification (Sentiment: Positive/Negative).
• Basic summarization.
• Translation of common phrases.

Example: "Review: 'The steak was overcooked.' Sentiment:" $\rightarrow$ Result: "Negative"

---

2. Few-Shot Prompting (The Industry Standard)

Few-Shot involves giving the model a small group of examples (usually 2 to 5) of the task being performed correctly. This is the most effective way to ensure a model follows a specific format or tone.

Why use it?

It dramatically reduces "Variation." If you want the model to extract names in a specific custom format like [NAME] - [AGE], a Zero-shot prompt will likely fail 20% of the time. A Few-shot prompt will succeed 99% of the time.

Task: Extract names and ages.

Example 1: Input: "Bob is 45." Output: [BOB] - [45]
Example 2: Input: "Alice just turned 30." Output: [ALICE] - [30]

Input: "David is 50." Output:

---

3. Chain-of-Thought (CoT): "Show Your Work"

For complex math, logic, or multi-step reasoning, models often make "Stupid" mistakes by rushing to the final answer. Chain-of-Thought forces the model to document its reasoning steps before giving the final result.

The "Let's Think Step-by-Step" Hook

By simply adding the phrase "Let's think step by step" to a prompt, you trigger the model's internal logic gate to process the intermediate steps.

graph LR
    A[Hard Question] --> B[Standard Prompt]
    B --> C[Wrong Answer: Rushed logic]
    
    A --> D[CoT Prompt]
    D --> E[Step 1: Reasoning]
    E --> F[Step 2: Logic Check]
    F --> G[Correct Answer: Verified logic]

Advanced: Manual Chain-of-Thought

As an LLM Engineer, you don't just rely on the model to "think step by step." You force the structure.

Professional Prompt Snippet: *"Perform the following steps:

1. Identify the core mathematical constraint.
2. Calculate the intermediate value.
3. Compare the intermediate value to the budget.
4. Provide the final recommendation."*

---

4. When to Use Which Technique?

Technique	Intelligence Needed	Latency/Cost	Best Use Case
Zero-Shot	Low	Low	General tasks, broad summaries.
Few-Shot	Medium	Medium (more tokens!)	Consistent formatting, specific tone.
Chain-of-Thought	High	High (Slow generation)	Math, Strategy, Coding, Legal.

---

Code Concept: Automated Few-Shot with Python

In a production app, you can't manually type examples every time. You store them in a list and inject them into a template.

examples = [
    {"input": "The weather is nice.", "output": "POSITIVE"},
    {"input": "I am so angry.", "output": "NEGATIVE"}
]

def format_few_shot_prompt(new_input):
    base_prompt = "Classify sentiment.\n\n"
    for ex in examples:
        base_prompt += f"Input: {ex['input']}\nOutput: {ex['output']}\n"
    
    base_prompt += f"Input: {new_input}\nOutput:"
    return base_prompt

print(format_few_shot_prompt("I love this course!"))

---

Summary

• Zero-Shot: Quick and dirty. Good for base intelligence.
• Few-Shot: Prodigious for formatting and consistency. Inject 3-5 high-quality examples.
• Chain-of-Thought: Slow but smart. Use for any task that requires logic or "checking" before answering.

In the next lesson, we will look at System Prompts and Personas, where we move from the "Body" of the prompt to the "Soul" of the agent.

---

Exercise: The Reasoning Challenge

Take a simple logic puzzle (e.g., "A father is 3 times older than his son. In 10 years, he will be twice as old. How old is the son now?").

1. Try it with a Zero-Shot prompt. Does it get it right?
2. Try it with a Chain-of-Thought prompt ("Think step by step").

Compare the responses. You will notice the CoT response is longer, but significantly more reliable.