Few-shot learning
A prompting technique where the LLM sees a small number of input/output examples in the prompt before being asked to perform the same task on a new input.
Few-shot learning is the most reliable prompt-engineering technique. Show the model 2–8 examples of the task; it generalizes the pattern and applies it to the next input. No fine-tuning, no labeled dataset — just well-chosen examples in the prompt.
The trick is example selection. Hand-curated examples that cover the diversity of expected inputs beat randomly sampled ones almost every time. For high-volume tasks, dynamic few-shot retrieval (find the K most similar past examples and inject them) often matches fine-tune-level accuracy.
In 2026, few-shot remains the default for any task where you have examples but not enough to fine-tune. Reasoning models reduce its impact on hard problems (they already reason internally), but for narrow classification, extraction, and style-matching tasks, few-shot is still the workhorse.
Frequently asked
How many examples should I include in a few-shot prompt?+
3–5 well-chosen examples usually beats 10+ random ones. For diverse tasks, dynamically retrieve the closest K examples to the current input rather than using a fixed set.
When should I fine-tune instead of using few-shot?+
When you have 1000+ labeled examples, when context-window cost dominates your token spend, or when accuracy plateaus despite well-tuned prompts. Otherwise few-shot is cheaper and easier to iterate on.