When it works
Concrete behaviors: 'Do not include headers.' 'Do not use bullet points.' Model can pattern-match negations.
Advertisement
When it backfires
Vague negatives ('Do not be biased') don't help. Naming forbidden content ('do not mention Coca-Cola') can INCREASE mention rate.
Advertisement
Positive framing
'Respond in a single paragraph' beats 'Do not use bullets.' State what you want, not what you don't.
Constraint list
End of prompt: bulleted constraints. Model attends to these. Order by importance (recency bias).