Positive prompts tell the model what to do. Negative prompts tell it what to avoid. Both are necessary: positive instructions set direction, negative instructions act as guardrails against common failure modes.
The anti-pattern is writing "don't X" for every possible failure (the list becomes longer than the actual task). Good negative prompting targets the 2-3 specific failure modes the model keeps producing. Combined with a positive instruction, it converges fast on the desired behavior.
Example Prompt
Summarize this article in 3 sentences.
Do NOT:
- Use the phrase "in today's fast-paced world" or any similar cliche opener
- Start with "The article discusses..."
- Include any information not present in the article
- Use bullet points or markdown
Article: [...]When to use it
- The model keeps making a specific wrong move (generic openers, extra sections, wrong format)
- You want to prevent a known failure pattern rather than describe everything right
- Pair with positive instructions for symmetric coverage
When NOT to use it
- You have no observed failure modes yet -- write positive instructions first
- The negative list grows past 5-6 items -- refactor to positive instructions instead
- You're using negative prompts as a workaround for a bad task definition