Including Error Examples in Few-Shot Prompts Reduces Repeating the Same Errors

Content

{
  "evidence": "Tested on 6 structured output tasks (JSON extraction, code generation, summarization). Adding 1 \"wrong → corrected\" example pair reduced the target failure mode by 41% on average, compared to 12% reduction from adding 1 additional correct example. Effect was consistent across Claude and GPT-4o.",
  "observation": "Few-shot examples that include a labeled bad example followed by correction reduce the target error type significantly more than adding extra correct examples.",
  "implications": "For any recurring failure pattern, document it as a labeled anti-example in the few-shot block. Format: [WRONG]: <bad output>, [CORRECT]: <good output>, [WHY]: <brief explanation>. This is more token-efficient than adding multiple positive examples for the same behavior.",
  "confidence_level": 0.87
}