Few-Shot vs Zero-Shot Prompting
Zero-shot asks a model to do a task from instructions alone; few-shot adds a handful of worked examples so the model can pattern-match instead of just following a description.
| Dimension | Few-Shot Prompting | Zero-Shot Prompting |
|---|---|---|
| Prompt length and token cost | Every example adds recurring input tokens, paid on every request. | Wins here No examples to pay for — just the instruction and the input. |
| Nailing an exact output format | Wins here Examples pin down JSON keys, label vocabulary, and edge-case handling precisely. | A written instruction alone often leaves format ambiguous at the edges. |
| Tasks with strong pretraining priors (translate, summarize, classify sentiment) | Examples add little when the model has already seen enormous numbers of this exact task. | Wins here Rides on pretraining alone, which is usually already well-calibrated for common tasks. |
| Bespoke, non-obvious rubrics (a custom taxonomy, company-specific labels) | Wins here Examples anchor a distinction the model has no pretrained prior for at all. | The model has to guess where the line sits, and different phrasings of the same case can land in different buckets. |
| Consistency across reruns | Wins here Calibrated examples reduce output variance on ambiguous cases. | Can bounce between plausible answers run to run when the task is genuinely ambiguous. |
| Setup effort | Choosing diverse, representative examples is itself a small design task. | Wins here Nothing to curate beyond the instruction itself. |
| Latency | A longer prompt takes marginally longer to process before generation starts. | Wins here Shorter prompt, slightly less to process upfront. |
When to choose Few-Shot Prompting
Add few-shot examples when output format or judgment is precise or idiosyncratic and hard to describe in words alone, or a scored eval shows zero-shot output is inconsistent.
When to choose Zero-Shot Prompting
Stay zero-shot when the task is general and well-represented in pretraining, or every token of prompt budget genuinely counts.
The verdict
Start zero-shot, score it against a fixed eval set, and add few-shot examples only where the evidence shows a specific, fixable gap — not as a reflex applied to every prompt.