Few-shot prompting "programs" the model by demonstration. Each example sets precedent: format, tone, depth, edge-case handling. The model generalizes from the examples to your real input.
The sweet spot is usually 2-4 examples. More than 5 adds cost without improving quality on modern models, and can actually hurt if the examples reinforce bias or contradict each other.
Example Prompt
Convert natural language to SQL.
Input: Show me customers who signed up last month.
Output: SELECT * FROM customers WHERE created_at >= date_trunc('month', now() - interval '1 month') AND created_at < date_trunc('month', now());
Input: Which products have more than 100 orders?
Output: SELECT product_id, COUNT(*) FROM orders GROUP BY product_id HAVING COUNT(*) > 100;
Input: How many users are in Texas?
Output:When to use it
- Output has a specific shape or style zero-shot doesn't capture
- You have high-quality example pairs readily available
- Task is narrow and consistent across inputs
- Using a smaller model where zero-shot underperforms
When NOT to use it
- Examples might bias the output away from the real input's context
- The task is trivial for the model (wasted tokens)
- Your examples contradict each other (worse than zero-shot)