🧠 AI with Python – 🎯 Zero-shot vs Few-shot Prompting

One of the biggest misconceptions about Large Language Models (LLMs) is that they're simply "smart enough" to answer any question you ask. While modern models like GPT, Gemma, Llama, and Mistral are incredibly capable, the quality of their responses depends heavily on how you ask the question.

This is where Prompt Engineering comes in.

Prompt Engineering is the process of designing effective prompts that guide an AI model toward producing accurate, relevant, and well-structured responses. A small change in wording can completely change the quality of the output, making prompt engineering one of the most valuable skills for anyone working with LLMs.

Among the many prompting techniques available, Zero-shot Prompting and Few-shot Prompting are two of the most fundamental. Understanding the difference between them is the first step toward building reliable AI applications.

---

What Is Zero-shot Prompting?

Zero-shot prompting is exactly what its name suggests—you ask the model to perform a task without giving it any examples.

Instead, you simply describe what you want.

For example:

Translate the following sentence into French.

Sentence:
Python is an amazing programming language.

Here, the model receives only the instruction. It relies entirely on the knowledge it learned during pretraining and instruction tuning to generate the correct response.

This approach works surprisingly well because modern LLMs have already been trained on massive amounts of text covering translation, summarization, question answering, reasoning, coding, and countless other tasks.

For many everyday applications, Zero-shot prompting is often all you need.

---

What Is Few-shot Prompting?

Few-shot prompting takes a different approach.

Instead of asking the model to perform a task immediately, you first provide a few examples that demonstrate exactly what kind of output you expect.

For example:

English: Good Morning
French: Bonjour

English: Thank You
French: Merci

English: Good Night
French: Bonne Nuit

English: Python is an amazing programming language.
French:

Notice what changed.

We never explicitly explain how to translate English into French. Instead, we simply show the model a few examples and allow it to recognize the pattern.

Large Language Models are remarkably good at pattern recognition. After seeing just a handful of examples, they can often continue the pattern with impressive accuracy.

This ability is known as In-Context Learning, one of the defining characteristics of modern LLMs.

---

Why Do Examples Matter?

Imagine you're training a new employee.

You could simply say:

"Write customer support replies."

Or you could provide three well-written examples before asking them to respond to a customer.

Most people would perform better after seeing examples.

Large Language Models behave in a very similar way.

Examples remove ambiguity by showing:

• the expected writing style,
• the desired output format,
• the level of detail,
• and even the tone of the response.

Instead of guessing what you want, the model learns directly from the examples you've included in the prompt.

---

Comparing Zero-shot and Few-shot Prompting

Suppose we want to classify movie reviews as either Positive or Negative.

With Zero-shot prompting, we might simply ask:

Classify the sentiment:

"I absolutely loved this movie."

Most modern LLMs will correctly identify the sentiment as Positive because sentiment analysis is a task they've already encountered during training.

Now consider the same task using Few-shot prompting:

Review: This phone is fantastic.
Sentiment: Positive

Review: The service was terrible.
Sentiment: Negative

Review: I absolutely loved this movie.
Sentiment:

The additional examples immediately establish a pattern.

Rather than deciding how the answer should look, the model simply follows the format you've demonstrated.

The result is usually more consistent, especially when processing hundreds or thousands of similar requests.

---

When Should You Use Zero-shot Prompting?

Zero-shot prompting works best when the task is straightforward and the model already understands it well.

Common examples include:

• answering general questions,
• translating between languages,
• summarizing articles,
• explaining concepts,
• generating code,
• brainstorming ideas,
• writing emails,
• and creating blog content.

Since no examples are required, prompts remain short, clean, and inexpensive to process.

For many real-world AI applications, Zero-shot prompting is the fastest and simplest solution.

---

When Is Few-shot Prompting Better?

Few-shot prompting becomes particularly valuable whenever consistency is important.

Suppose you're extracting structured information from invoices.

Without examples, the model might return:

Invoice Number: INV-1001

or

Number = INV-1001

or

{
    "invoice": "INV-1001"
}

All three responses are technically correct—but they aren't consistent.

If your application expects the same format every time, inconsistent outputs create additional work.

By including just a few examples, you teach the model exactly how every response should be structured.

This is why Few-shot prompting is widely used for:

• document extraction,
• information classification,
• structured JSON generation,
• Named Entity Recognition,
• customer support automation,
• report generation,
• and enterprise AI workflows.

---

Zero-shot vs Few-shot: Which One Is Better?

Neither technique is universally better.

Instead, they solve different problems.

Zero-shot Prompting

✅ Simple to write

✅ Faster prompts

✅ Lower token usage

✅ Ideal for general-purpose tasks

Few-shot Prompting

✅ More consistent responses

✅ Better formatting

✅ Improved performance on specialised tasks

✅ Excellent for structured outputs

A good rule of thumb is to start with Zero-shot prompting. If the responses aren't consistent enough for your application, introduce a few carefully chosen examples and switch to Few-shot prompting.

---

Prompt Engineering Is More Than Asking Questions

Many beginners think Prompt Engineering is simply "asking AI better questions."

In reality, it's about designing instructions.

A well-crafted prompt can influence:

• the reasoning process,
• the writing style,
• the response format,
• the level of detail,
• the creativity,
• and even how confidently the model answers.

This is why Prompt Engineering has become one of the most valuable skills in modern AI development.

Whether you're building chatbots, AI coding assistants, customer support systems, or Retrieval-Augmented Generation (RAG) applications, the quality of your prompts directly impacts the quality of your results.

---

Final Thoughts

Zero-shot and Few-shot Prompting are the foundation of Prompt Engineering. While Zero-shot prompting leverages the model's existing knowledge with simple instructions, Few-shot prompting goes a step further by teaching the model through examples. Understanding when to use each technique allows you to build AI applications that are not only more accurate but also more consistent and reliable.

In this article, we explored how both prompting strategies work, why examples improve model performance, and where each approach is most effective. As you continue your journey into Prompt Engineering, you'll discover that writing effective prompts is just as important as choosing the right model.

---

from transformers import pipeline

# =========================================================
# Load Instruction-Following Model
# =========================================================

print("=" * 70)
print("Loading Instruction-Following Language Model...")
print("=" * 70)

generator = pipeline(
    "text-generation",
    model="google/gemma-2b-it"
)

# =========================================================
# Zero-shot Prompting
# =========================================================

zero_shot_prompt = """
Translate the following sentence into French.

Sentence:
Python is an amazing programming language.
"""

print("\n" + "=" * 70)
print("Zero-shot Prompt")
print("=" * 70)
print(zero_shot_prompt.strip())

zero_shot_response = generator(
    zero_shot_prompt,
    max_new_tokens=60,
    do_sample=False
)

zero_shot_output = zero_shot_response[0]["generated_text"]

print("\nZero-shot Response:")
print("-" * 70)
print(zero_shot_output)

# =========================================================
# Few-shot Prompting
# =========================================================

few_shot_prompt = """
English: Good Morning
French: Bonjour

English: Thank You
French: Merci

English: Good Night
French: Bonne Nuit

English: Python is an amazing programming language.
French:
"""

print("\n" + "=" * 70)
print("Few-shot Prompt")
print("=" * 70)
print(few_shot_prompt.strip())

few_shot_response = generator(
    few_shot_prompt,
    max_new_tokens=60,
    do_sample=False
)

few_shot_output = few_shot_response[0]["generated_text"]

print("\nFew-shot Response:")
print("-" * 70)
print(few_shot_output)

# =========================================================
# Compare Zero-shot and Few-shot Outputs
# =========================================================

print("\n" + "=" * 70)
print("Zero-shot vs Few-shot Comparison")
print("=" * 70)

print("\nZero-shot Output:")
print(zero_shot_output)

print("\nFew-shot Output:")
print(few_shot_output)

# =========================================================
# Zero-shot Prompting for Multiple Tasks
# =========================================================

zero_shot_tasks = [
    "Classify the sentiment as Positive or Negative: I absolutely love this movie.",
    "Summarize in one sentence: Python is a high-level programming language designed for readability and rapid development.",
    "Explain recursion in one simple sentence."
]

print("\n" + "=" * 70)
print("Zero-shot Prompting for Multiple Tasks")
print("=" * 70)

for index, task in enumerate(zero_shot_tasks, start=1):
    print(f"\nTask #{index}")
    print("-" * 70)
    print(task)

    response = generator(
        task,
        max_new_tokens=80,
        do_sample=False
    )

    output = response[0]["generated_text"]

    print("\nResponse:")
    print(output)

# =========================================================
# Few-shot Prompting for Sentiment Classification
# =========================================================

classification_prompt = """
Review: This phone is fantastic.
Sentiment: Positive

Review: The service was terrible.
Sentiment: Negative

Review: The product is acceptable but not impressive.
Sentiment: Neutral

Review: I am extremely happy with this purchase.
Sentiment:
"""

print("\n" + "=" * 70)
print("Few-shot Sentiment Classification")
print("=" * 70)
print(classification_prompt.strip())

classification_response = generator(
    classification_prompt,
    max_new_tokens=20,
    do_sample=False
)

classification_output = classification_response[0]["generated_text"]

print("\nClassification Response:")
print("-" * 70)
print(classification_output)

# =========================================================
# Few-shot Prompting for Structured Output
# =========================================================

structured_prompt = """
Text: Python was created by Guido van Rossum.
Output: {"language": "Python", "creator": "Guido van Rossum"}

Text: Java was created by James Gosling.
Output: {"language": "Java", "creator": "James Gosling"}

Text: Ruby was created by Yukihiro Matsumoto.
Output:
"""

print("\n" + "=" * 70)
print("Few-shot Structured Output")
print("=" * 70)
print(structured_prompt.strip())

structured_response = generator(
    structured_prompt,
    max_new_tokens=40,
    do_sample=False
)

structured_output = structured_response[0]["generated_text"]

print("\nStructured Response:")
print("-" * 70)
print(structured_output)

# =========================================================
# Few-shot Prompting for Text Classification
# =========================================================

topic_prompt = """
Text: The team won the football championship.
Category: Sports

Text: The company released a new smartphone.
Category: Technology

Text: The government announced a new tax policy.
Category: Politics

Text: Researchers developed a faster machine learning model.
Category:
"""

print("\n" + "=" * 70)
print("Few-shot Topic Classification")
print("=" * 70)
print(topic_prompt.strip())

topic_response = generator(
    topic_prompt,
    max_new_tokens=20,
    do_sample=False
)

topic_output = topic_response[0]["generated_text"]

print("\nTopic Classification Response:")
print("-" * 70)
print(topic_output)

# =========================================================
# Reusable Prompt Generation Function
# =========================================================

def generate_response(prompt, max_new_tokens=80):
    response = generator(
        prompt,
        max_new_tokens=max_new_tokens,
        do_sample=False
    )
    return response[0]["generated_text"]


# =========================================================
# Compare Both Techniques Using the Same Task
# =========================================================

comparison_zero_shot = """
Classify the following customer message as Complaint, Question, or Praise.

Message:
The support team resolved my issue very quickly.
"""

comparison_few_shot = """
Message: My order arrived damaged.
Category: Complaint

Message: When will my subscription renew?
Category: Question

Message: Your support team was incredibly helpful.
Category: Praise

Message: The support team resolved my issue very quickly.
Category:
"""

print("\n" + "=" * 70)
print("Same Task: Zero-shot vs Few-shot")
print("=" * 70)

print("\nZero-shot Result:")
print("-" * 70)
print(generate_response(comparison_zero_shot, max_new_tokens=30))

print("\nFew-shot Result:")
print("-" * 70)
print(generate_response(comparison_few_shot, max_new_tokens=30))

# =========================================================
# Program Completed
# =========================================================

print("\n" + "=" * 70)
print("Zero-shot vs Few-shot Prompting Demo Completed Successfully!")
print("=" * 70)