Zero-shot & few-shot

The cheapest way to specify a task is to describe it (zero-shot); the next is to show it with examples (few-shot), exploiting the model's in-context learning. Knowing which to reach for saves tokens and improves reliability.

Zero-shot: just ask

A clear instruction with no examples. Modern instruct models are strong zero-shot for common tasks. Start here — it's the cheapest and often enough.

Few-shot: show the pattern

Include a handful of input → output demonstrations. Use it when:

• The task is easier to show than to describe (a specific format, a nuanced labeling rule, a tone).
• You need output consistency — examples pin down the exact shape.
• Zero-shot is close but unreliable on edge cases.

Doing few-shot well

• Match the distribution — examples should resemble real inputs, including tricky cases; the model imitates what it sees.
• Be consistent — identical format across examples; the format is the instruction.
• Cover the classes — for classification, include each label (and watch ordering; models can pick up position bias).
• 2–5 is usually enough — more costs tokens and context for diminishing returns; for many tasks one good example beats three mediocre ones.

Pitfall

Few-shot examples can over-anchor the model to their surface form (it copies a quirk you didn't intend). And they consume context budget on every call — for high-volume tasks, consider whether fine-tuning beats permanent few-shot.

Connects to: in-context learning · chain-of-thought · few-shot vs fine-tune