5 min

Zero-Shot & Few-Shot

Just Ask vs. Show Examples

Zero-Shot: Just Ask

The simplest prompting technique: describe what you want and let the model figure it out. No examples, no special formatting — just a clear instruction.

Classify this email as "urgent", "normal", or "low-priority":

Subject: Server down - all production APIs failing
Body: Our main API server went offline at 3:42 AM. Customer-facing services are returning 500 errors. Engineering is investigating.

Zero-shot works well when:

The task is common (summarization, classification, translation)

The categories are self-explanatory

You don't need a specific output format

Zero-shot fails when:

The task requires domain-specific knowledge ("classify by our internal ticket categories")

The output format matters ("return exactly this JSON schema")

Edge cases exist that the model might not anticipate

Few-Shot: Show, Don't Tell

Instead of explaining what you want, show examples of input → output pairs. The model pattern-matches against your examples.

Classify these emails by department:

Email: "Can I upgrade my plan to enterprise?"
Department: Sales

Email: "The export button isn't working in Chrome"
Department: Support

Email: "We'd like to schedule a demo for our team"
Department: Sales

Email: "Invoice #4521 has the wrong billing address"
Department: ???

The model sees the pattern and predicts "Billing" or "Finance" — without you ever defining what departments exist.

How Many Examples?

Count	When to Use
1-2	Simple tasks, clear categories
3-5	Ambiguous tasks, subtle distinctions
5-10	Complex formatting, domain-specific output

More examples aren't always better. After 5-7 examples, you get diminishing returns and waste context window space.

Few-Shot for Extraction

Extract contact information from these emails:

Email: "Hi, I'm Sarah Chen from Acme Corp. Reach me at sarah@acme.com or 555-0123."
Result: {"name": "Sarah Chen", "company": "Acme Corp", "email": "sarah@acme.com", "phone": "555-0123"}

Email: "This is Mike at GlobalTech (mike.r@globaltech.io). Our billing address is 123 Main St."
Result: {"name": "Mike", "company": "GlobalTech", "email": "mike.r@globaltech.io", "phone": null}

Email: "Please contact our procurement team at procurement@wayne.co for next steps."
Result: ???

Notice the second example includes a null for phone — this teaches the model how to handle missing fields.

Few-Shot for Generation

Write a one-line product tagline in our brand voice:

Product: Project management tool for remote teams
Tagline: "Ship projects, not status updates."

Product: AI-powered email assistant
Tagline: "Your inbox, minus the busywork."

Product: Cybersecurity training platform
Tagline: ???

The examples establish tone, length, and style — the model continues the pattern.

Choosing Between Zero-Shot and Few-Shot

Scenario	Use
Well-known task, flexible output	Zero-shot
Custom categories or labels	Few-shot
Specific output format required	Few-shot
Quick prototype or exploration	Zero-shot
Production pipeline with consistency needs	Few-shot

Practice Tasks

Try these with the sample data in your project:

Zero-shot classification: Classify the emails in data/emails.json as urgent/normal/low-priority

Few-shot extraction: Extract sender name, company, and intent from each email using 2-3 examples

Few-shot generation: Write professional reply suggestions using example replies as style guides

Key Takeaways

Zero-shot is fast and works for common tasks. Start here, then add examples if quality is low.

Few-shot teaches by example — great for custom formats, domain labels, and consistent output.

Include edge cases in your examples (nulls, ambiguous inputs) to handle them gracefully.

3-5 examples is the sweet spot for most tasks.

This is chapter 2 of Prompt Engineering Essentials.

Get the full hands-on course — free during early access. Build the complete system. Your projects become your portfolio.

View course details

Ch. 1: How LLMs Think

Ch. 3: Chain-of-Thought