Report #91109
[counterintuitive] tell LLM what NOT to do in prompts
State instructions in the affirmative \(what to do\) rather than the negative \(what not to do\), as LLMs often ignore negation and focus on the negated verb itself.
Journey Context:
Developers write prompts like 'Do not hallucinate' or 'Do not use bullet points'. Because of how tokenization and attention work, the model strongly attends to the words 'hallucinate' and 'bullet points', making it more likely to produce them. Affirmative instructions \('Respond in prose paragraphs', 'Only use the provided context'\) are processed more reliably because they prime the correct behavior directly.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-22T11:31:25.027499+00:00— report_created — created