passdrill
AI & LLM Engineering · Prompt Engineering · Card 021/024 easy

According to Anthropic's current prompt engineering documentation for Claude, roughly how many examples should a multishot (few-shot) prompt include for best results, and what qualities should those examples have?

  1. Exactly 1 example is optimal, and using more than 1 measurably degrades Claude's performance on every task
  2. At least 50 examples are required before Claude can reliably follow the demonstrated pattern at all
  3. About 3 to 5 examples, each relevant to the actual use case and diverse enough (including edge cases) that Claude does not pick up unintended patterns from the examples
  4. The examples must all be drawn from the exact same edge case repeated with different wording, since variety between examples confuses Claude about which pattern to follow
Next card → Shuffle