Zero-Shot vs Few-Shot: When Extra Examples Help
Choose zero-shot for clear, well-known tasks; add few-shot examples when format or reasoning is unusual. Learn where to put instructions in long context—and how to A/B both approaches in one chat.
Large language models can follow a short instruction (zero-shot) or learn a pattern from a few input→output examples in the same prompt (few-shot). The useful question is not which technique is “smarter”—it is when examples pay for the extra tokens.
When zero-shot is enough
Use a clear instruction and skip examples when the task is:
- Well-defined — classify into known labels, extract obvious fields, summarize with a stated length
- Covered by common training patterns — rewrite tone, translate, answer a factual question from attached text
- Schema-simple — “return JSON with keys X, Y, Z” and the rules are explicit
Extra examples here mostly burn context and can make the model copy an odd style from your samples.
When few-shot (or CoT) pays off
Add 1–3 short exemplars when you need:
- A custom format that is hard to describe in words alone
- Domain nuance (edge cases, house style, unusual rubrics)
- Multi-step reasoning — show the steps once, then ask for the same process on new input
A practical pattern:
[Instruction]
[Example 1: input → brief reasoning → output]
[Example 2: input → brief reasoning → output]
[Your real input]
Published model comparisons often report different strengths by benchmark (for example GPQA and MMLU figures that vary by model and date). Treat those numbers as snapshots, not permanent rankings—always re-check on your task.
Long context: where instructions go
Very long prompts can bury critical rules in the middle (“lost in the middle”). Put:
- Must-follow rules and output schema at the top
- Source material in the middle
- A short reminder of the format at the end
Use markdown headings or simple tags (<constraints>, <source>) so the model can anchor attention.
Quick decision table
| Situation | Prefer |
|---|---|
| Clear labels / simple extract | Zero-shot |
| Odd schema or house style | Few-shot |
| Hard math / multi-step logic | Few-shot + step-by-step |
| Huge document + short ask | Zero-shot; put rules at start and end |
| Comparing models on same task | Same prompt; swap models |
Try both on AI Plaza
On AI Plaza Chat:
- Write one prompt (zero-shot).
- Duplicate it and insert two short examples (few-shot).
- Keep Auto, or lock the same model for a fair A/B.
- For repeat formats, save the pattern as a scenario tool instead of retyping examples.
See also: Prompt engineering basics and How to use prompt tools.