Module 1 · Designing prompts systematically · Lesson 2
Choosing few-shot examples
What examples actually teach a model, when they help and when they anchor it, how to select, order and place them (static, retrieved, many-shot), and how to measure whether they earn their tokens, using a follow-up questions drafter.
By the end of this lesson you can
- Explain what in-context examples teach (format, label space, input distribution) and why that makes them powerful and risky
- Choose between no examples, a static curated set, retrieved examples and many-shot prompting for a given step, and justify the choice
- Implement example selection with similarity plus diversity (MMR) and stratification, keeping the example pool disjoint from the eval set
- Place examples so prompt caching still works, and compute the cost difference between static and dynamic examples
- Design an ablation that measures whether examples help, including an anchoring check