Lesson 3 · 6 min
Few-Shot Examples
How many examples, in what order, of what shape.
3–5 examples typically peaks. Make them diverse and structurally identical to the desired output. Put the hardest example last — recency bias means Claude weights it most.
Production scenario
Real-world example: Email triage at a customer-success team
A CS team auto-routes inbound mail. The triage prompt uses five carefully chosen examples:
- Newsletter (easy)
- Refund request (common case)
- Partnership pitch (edge case — looks like a customer email but isn't)
- Urgent outage report (the high-stakes one)
- Ambiguous "is this an upsell?" (the hardest)
The hardest is last. Within a week, the accuracy on the ambiguous bucket climbs 11 points — recency bias is doing the work.
Why this matters: three to five diverse examples, hardest last. Don't pack twenty similar examples; the model overfits to the shape, not the judgment.
Knowledge points in this lesson
- 3–5 examples is the typical sweet spot
- Examples should be structurally identical
- Put hardest example last (recency bias)
- Diverse examples beat similar ones
- Too many examples can confuse the model
Quick check
Prompt EngineeringSelect all that apply
Which of these maximize prompt cache hit rate? (Select all that apply)