Why it matters
Many tasks that fail zero-shot succeed with 3-5 well-chosen examples. Few-shot is cheap (no training) and reversible (change prompt any time). Understanding it well is core prompt engineering.
The architecture
Structure: instruction + examples (input → output pairs) + new input. Model sees the pattern and applies it.
Number of examples: 3-5 typical. Diminishing returns beyond ~10. Very long few-shot uses tokens without proportional gain.
How it works end to end
Example selection: pick examples that cover the task's range. Include tricky cases you want the model to handle. Diversity of examples > raw count.
Format consistency: examples must use the exact format you want in output. Model imitates the pattern precisely.
Dynamic few-shot: for classification, retrieve most similar training examples per query. Higher quality than fixed examples.