Stop 12 · Prompt EngineeringTopics · 2/10

Zero-Shot & Few-Shot Prompting

In short

Sometimes just asking works. Sometimes the model needs to see two or three examples first. Knowing when to add examples saves you from fine-tuning for simple jobs.

Zero-shot asks a model to perform a task with no examples; few-shot includes a handful of examples in the prompt to steer format and style — a cheap alternative to fine-tuning.

The blob shows three sample mehendi cards to an artist who draws the same style on a palm.

Think of it like this

Showing the mehendi artist samples

Tell a mehendi artist 'something nice' and she'll draw her usual design. Show her three sample cards of the exact style you want, and she copies that pattern on your palm. Zero-shot is just asking. Few-shot is showing samples first.

Watch it happen

No examples vs a few examples

Zero-shot

Just the instruction
Short and cheap
Format may vary

Few-shot

Instruction plus 2 to 5 examples
Locks in format and tone
Costs more tokens

Start zero-shot. If answers drift in format, labels or tone, add a few short, varied examples that show exactly what you want.

The big ideas

What you need to know

Zero-shot for common tasks

Summarise, translate, classify into obvious groups. Strong models handle these with no examples.

Few-shot teaches the pattern

Examples show the exact labels, length and style. The model copies the pattern closely.

Vary your examples

If all examples are positive, the model leans positive. Mix classes and include an edge case.

Examples leak

Models sometimes copy example details into answers. Use clearly different examples from the real input.

See it in code

Walk through the code

text

1Classify each customer message as ORDER, REFUND or OTHER.2 3Message: "Where is my parcel? It said delivered today."4Label: ORDER5 6Message: "I got the wrong size, I want my money back."7Label: REFUND8 9Message: "Do you have a store in Pune?"10Label: OTHER11 12Message: "Paid twice by UPI for the same order, please return one."13Label:

1/3The instruction alone would be zero-shot.

Three examples fix the label names and the one-word answer format. The last line leaves a blank for the model to fill.

Save yourself the pain

Common mistakes

Oops

Adding 30 examples when 3 would do.

Do this instead

More examples cost tokens and slow responses. Add examples only until the output is consistent.

Oops

Using examples that are all the same class.

Do this instead

Balance the classes. Otherwise the model learns 'the answer is usually X'.

Your mission

Label 10 messages

~25 min

Quick check

Your zero-shot classifier sometimes answers 'Refund request' and sometimes 'REFUND'. What's the simplest fix?

Go deeper

1Compare zero-shot vs. few-shot output quality on the same task.
2Practice picking diverse, representative few-shot examples.
3Learn when few-shot stops helping and fine-tuning becomes worth the investment.