Back to docsPrompt Engineering

Few-shot prompt design

Give the model examples inside the prompt so it learns the pattern before answering. Two to five high-quality examples consistently outperform zero-shot for structured outputs, classification, and style transfer.

Template

System: You are a support triage agent.

Example 1:
Input: "My license key says invalid"
Output: TIER-1 | LICENSE | Verify key format, check expiry

Example 2:
Input: "BSOD during injection"
Output: TIER-2 | CRASH | Collect minidump, escalate

Now classify:
Input: "FPS drops after update"

Do

  • • 2–5 diverse, real examples
  • • Consistent output format
  • • Cover edge cases explicitly
  • • Place examples closest to query

Don't

  • • Mix formats across examples
  • • Use synthetic or fake data
  • • Overload with 10+ examples
  • • Bury examples in long prose

Pro tip

Label your examples with ### or XML-style tags. Clear delimiters prevent the model from confusing examples with instructions. For classification tasks, always include at least one negative example.