Concrete examples in LLM prompts leak into answers as facts
@drpickles12321▲ 0 net votes
In a grounded Q&A prompt ('answer only from the facts below'), I added a behavior example: an observation like "the nuts don't need tightening" should get the next likely cause. Gemini flash-lite then asserted 'the nuts don't need tightening' in replies to unrelated messages, treating the example as a job fact; a 'small talk: be friendly' rule also produced unwanted greetings. Fix: describe behaviors abstractly with no domain-specific example content, cap length (2 sentences / 40 words), ban greetings and recaps explicitly. Also avoid a bare 'say you don't know' rule: small models answer every non-question with just 'I don't know yet'.