Few-Shot Mastery: Teaching by Example

The most reliable way to get AI to produce exactly what you want: show it examples. Few-shot prompting is the technique of including 3-10 input/output examples in your prompt so the AI learns your desired pattern.

🔄 Quick Recall: In the previous lesson, you learned reasoning techniques (CoT, ToT, self-consistency) that improve how AI thinks. Few-shot prompting improves what the AI produces — by showing it the exact pattern you want.

Zero-Shot vs. Few-Shot

Zero-shot (no examples): “Classify this support ticket as Bug, Feature Request, or Question.”

Few-shot (with examples):

<examples>
Input: "The app crashes when I click export"
Category: Bug

Input: "Can you add dark mode?"
Category: Feature Request

Input: "How do I reset my password?"
Category: Question

Input: "Login page shows blank screen on Firefox"
Category: Bug

Input: "Is there a student discount?"
Category: Question
</examples>

<input>The search filter doesn't work when I select multiple tags</input>

The few-shot version is dramatically more reliable because the AI sees exactly what each category looks like — not just the category name.

The Art of Choosing Examples

Not all examples are equal. Strategically chosen examples outperform random ones.

Rule 1: Cover Every Category

For classification, include at least one example per category. Better: 2-3 per category.

Rule 2: Include Edge Cases

The most valuable examples are the ones that could go either way:

<examples>
<!-- Clear cases -->
Input: "I love this product! Best purchase ever!"
Sentiment: Positive

Input: "Terrible experience. I want a refund."
Sentiment: Negative

<!-- Edge cases — these teach the AI WHERE the boundary is -->
Input: "The product works fine, but the packaging was damaged."
Sentiment: Mixed

Input: "Not bad, I guess. Does what it says."
Sentiment: Neutral

Input: "Great concept, but the execution needs work."
Sentiment: Mixed
</examples>

Without edge cases, the AI handles clear-cut inputs well but guesses on ambiguous ones. Edge cases teach the decision boundaries.

Rule 3: Vary Input Characteristics

If all your examples have similar length, style, or complexity, the AI may overfit to those characteristics:

<!-- BAD: All examples are short, simple sentences -->
Input: "I love it"     → Positive
Input: "I hate it"     → Negative
Input: "It's okay"     → Neutral

<!-- GOOD: Varied length, style, and complexity -->
Input: "I love it"
→ Positive

Input: "After using this for three months, I can confidently say
it has transformed how our team collaborates. The learning curve
was steep initially, but worth every minute."
→ Positive

Input: "meh. works but nothing special tbh"
→ Neutral

✅ Quick Check: You built a few-shot prompt for extracting meeting action items. Your 5 examples all show well-formatted meeting notes with clear “ACTION: Person does X by Date” entries. In practice, real meeting notes are messy — typos, unclear assignments, missing dates. Will your prompt work on real data? (Answer: Probably not well. Your examples only show the “clean” case. Add examples with messy input: “john said he’d look into the budget thing maybe next week” → Action: John to review budget by [next week]. Showing the AI how to handle messy, ambiguous input is what makes few-shot prompts work in production.)

Few-Shot for Different Tasks

Style Transfer

Teach the AI your writing voice:

<task>Write a product update email in our company's voice.</task>

<examples>
Topic: New dashboard feature
Our style: "Your dashboard just got smarter. We added real-time
alerts that ping you when metrics move — no more checking
manually. It's live now. Go poke around."

Topic: Pricing change
Our style: "Heads up: we're adjusting pricing starting March 1.
Current plans are locked in until your renewal. Here's what's
changing and why."

Topic: Maintenance window
Our style: "Quick note: we're doing some housekeeping this
Saturday 2-4 AM EST. Things might be spotty for a few minutes.
We'll be done before your coffee kicks in."
</examples>

<topic>New API rate limit increase</topic>

Three examples teach the AI your exact voice: casual, direct, first-person plural, short sentences, no marketing fluff.

Data Extraction

Teach the AI your extraction schema:

<task>Extract structured data from this invoice.</task>

<examples>
Input: "Invoice #1234 from Acme Corp, dated Jan 15 2026.
Items: Web hosting ($99/mo), SSL cert ($29/yr). Total: $128.
Payment due: Feb 15 2026."

Output:
{
  "invoice_number": "1234",
  "vendor": "Acme Corp",
  "date": "2026-01-15",
  "items": [
    {"description": "Web hosting", "amount": 99, "recurring": "monthly"},
    {"description": "SSL cert", "amount": 29, "recurring": "yearly"}
  ],
  "total": 128,
  "due_date": "2026-02-15"
}
</examples>

Code Generation

Teach the AI your coding patterns:

<task>Write a Python function following our conventions.</task>

<examples>
Request: "Function to validate email format"
Code:
def validate_email(email: str) -> bool:
    """Validate email format using regex pattern."""
    import re
    pattern = r'^[a-zA-Z0-9._%+-]+@[a-zA-Z0-9.-]+\.[a-zA-Z]{2,}$'
    return bool(re.match(pattern, email))
</examples>

<request>Function to calculate percentage change between two numbers</request>

How Many Examples?

Token budget matters: Each example consumes tokens from your context window. For a task with a 4K context, you have less room for examples than with 100K context. Prioritize example quality over quantity.

Practice Exercise

1. Create a few-shot classification prompt with at least 3 categories and 2 examples per category
2. Include one edge case per category that tests the boundary
3. Try the same task zero-shot vs. few-shot — compare consistency
4. Build a style transfer prompt: give 3 examples of your writing voice, then generate a new piece
5. Experiment with example order: move your best example to the end and see if output improves

Key Takeaways

• Few-shot prompting (3-10 examples) is the most reliable technique for consistent, patterned output
• Example quality matters more than quantity — diverse, edge-case-including examples outperform many similar ones
• Cover every category, include edge cases, and vary input characteristics to prevent overfitting
• Recency bias: place your most important examples last for maximum influence
• Few-shot works for classification, style transfer, data extraction, code generation, and more
• Token budget: balance example count with available context window size

Up Next

In the next lesson, you’ll learn to build system prompts and role definitions that shape AI behavior at a deeper level — creating reusable AI “agents” for specific tasks.