Skip to content
Home › Prompt Engineering & Structured Output

Few-Shot Prompting

Prompt Engineering & Structured Output

Definition

Multishot (few-shot) prompting supplies concrete input-output examples to steer Claude's output format, tone, and structure — "one of the most reliable ways to steer Claude's output."

Key points

  • Include 3–5 examples for best results.
  • Make examples Relevant (mirror the real use case), Diverse (cover edge cases; vary enough that Claude doesn't latch onto unintended patterns), and Structured (wrap in <example> tags, multiple in <examples>).
  • Few-shot works with thinking: put <thinking> tags inside the examples to demonstrate the reasoning pattern.
  • Fix inconsistent output with examples, not prose: "inconsistent output → add 3–4 few-shot examples, NOT more prose." Do not rely on temperature 0 alone for format normalization.
  • Format-normalization pattern (Architect's Playbook): provide 2–3 complete input-output examples with the standardized format (e.g. "cotton blend" / "Cotton/Polyester mix" → "Cotton Blend").
  • You can ask Claude to evaluate your examples for relevance/diversity or generate more from a seed set.

Why it matters for the exam

  • The "inconsistent output → few-shot examples not more prose" rule is a classic D4 distractor-killer.
  • Few-shot is the Playbook's mitigation for the Accuracy constraint (Structured Intermediates & Few-Shot Prompts).

Common gotchas

  • "Add more instructions / more prose" is the wrong fix for inconsistent formatting — add examples.
  • Too-similar examples cause Claude to overfit an unintended pattern; diversity matters.

See also

Sources

Referenced by