Few-Shot Prompting: Teach AI by Example in 5 Minutes

You don’t need to fine-tune a model to teach it new tricks. You just need good examples.

Few-shot prompting is the art of showing an AI what you want by giving it 2-5 examples, then letting it pattern-match on your actual task. It’s fast, flexible, and works shockingly well.

Zero-Shot vs Few-Shot vs Many-Shot

Zero-shot: No examples. Just instructions.

Classify this review as positive or negative: "The battery dies after 2 hours."

Few-shot: 2-5 examples before your actual task.

Review: "Best purchase I've made this year!" → Positive
Review: "Broke after one week." → Negative
Review: "It's okay, nothing special." → Neutral

Review: "The battery dies after 2 hours." → 

Many-shot: 10+ examples. Useful for nuanced tasks but eats context window.

For most tasks, 3-5 examples hit the sweet spot between accuracy and efficiency.

Why Few-Shot Works

Large language models are pattern completion engines. When you provide examples, you’re not “training” the model — you’re activating the relevant patterns it already learned during pre-training.

The examples serve three purposes:

1. Define the task without ambiguous instructions
2. Set the format by demonstration
3. Calibrate the difficulty and expected depth

The Golden Rules of Example Selection

Rule 1: Diversity Over Quantity

Three diverse examples beat ten similar ones.

Bad (all similar):

"I love this product!" → Positive
"Amazing quality!" → Positive  
"Best ever!" → Positive
"The screen cracked." → ?

Good (covers the spectrum):

"I love this product!" → Positive
"The screen cracked on day one." → Negative
"It works fine, nothing special." → Neutral
"The screen cracked." → ?

Rule 2: Include Edge Cases

If your task has tricky cases, show one in your examples.

"Not bad at all, actually pretty good." → Positive
"I wouldn't say it's terrible." → Neutral
"Could be worse, I guess." → Negative

Sarcasm, double negatives, and ambiguity are where few-shot really earns its keep.

Rule 3: Match Your Target Format Exactly

If you want JSON output, your examples should output JSON. If you want bullet points, your examples should use bullet points.

Input: "Python web framework"
Output: {"name": "Flask", "category": "web", "difficulty": "beginner"}

Input: "JavaScript runtime"
Output: {"name": "Node.js", "category": "runtime", "difficulty": "intermediate"}

Input: "Container orchestration"
Output:

The model will mirror your format precisely.

Practical Templates

Text Classification

Classify the support ticket priority.

"My account is hacked and money is missing" → P0-Critical
"The app crashes when I upload photos" → P1-High  
"Can you change the font size option?" → P3-Low
"Login page loads slowly on weekends" → P2-Medium

"I can't access any of my files since the update" →

Data Extraction

Extract structured data from the job posting.

Posting: "Senior React Developer at Stripe, NYC, $180-220k, 5+ years exp"
Result: {role: "Senior React Developer", company: "Stripe", location: "NYC", salary: "$180-220k", experience: "5+ years"}

Posting: "ML Engineer, Google, Remote US, $200-280k, PhD preferred"
Result: {role: "ML Engineer", company: "Google", location: "Remote US", salary: "$200-280k", experience: "PhD preferred"}

Posting: "DevOps Lead at Shopify, Toronto, CAD 150-190k, 7+ years"
Result:

Style Transfer

Rewrite the sentence in a casual, friendly tone.

Formal: "We regret to inform you that your application has been unsuccessful."
Casual: "Hey, unfortunately we won't be moving forward with your application this time."

Formal: "Please be advised that the system will undergo maintenance."
Casual: "Heads up — we're doing some maintenance on the system."

Formal: "Your inquiry has been received and will be processed within 5 business days."
Casual:

Code Generation

Write a Python function based on the description.

Description: "Check if a string is a palindrome"
Code:
def is_palindrome(s: str) -> bool:
    cleaned = s.lower().replace(" ", "")
    return cleaned == cleaned[::-1]

Description: "Find the most common element in a list"
Code:
def most_common(lst: list):
    from collections import Counter
    return Counter(lst).most_common(1)[0][0]

Description: "Merge two sorted lists into one sorted list"
Code:

Advanced: Chain Few-Shot with CoT

Combine few-shot examples with chain-of-thought reasoning for maximum accuracy:

Question: If a shirt costs $25 and is 20% off, what do you pay?
Thinking: Original price is $25. Discount is 20% of $25 = $5. Final price = $25 - $5 = $20.
Answer: $20

Question: A laptop is $800 with a 15% discount plus 8% tax on the discounted price. What's the total?
Thinking:

This gives the model both the pattern (few-shot) and the reasoning approach (CoT).

When Few-Shot Fails

1. Task is too novel. If the model has never seen anything like your task during training, examples alone won’t help. You might need fine-tuning.
2. Examples are misleading. Bad examples teach bad patterns. Garbage in, garbage out.
3. Context window overflow. Too many examples leave no room for the actual task. Keep it lean.

Recommended Gear

Key Takeaways

1. 3-5 diverse examples usually outperform lengthy instructions.
2. Your examples define the task, format, and difficulty — choose them carefully.
3. Always include at least one edge case.
4. Combine with CoT for complex reasoning tasks.
5. Test with your actual data, not just toy examples.

Few-shot prompting is the closest thing to a universal technique in prompt engineering. Master it, and you’ll handle 80% of tasks without any fancy tooling.