Negative Instructions
What you'll learn: How to use "do not" statements to steer model outputs away from unwanted behaviors while maintaining clarity.
The Core Idea
Negative instructions explicitly tell the model what not to do, rather than only specifying what you want. While positive instructions guide toward desired outcomes, negative instructions act as guardrails, preventing the model from taking unwanted paths during generation.
Think of it like giving directions: "Turn left at the intersection" is positive, but adding "Do not take the highway" prevents a common mistake. In prompting, this means statements like "Do not include personal opinions" or "Do not generate code examples."
Why Negative Instructions Matter
Large language models sometimes exhibit behaviors you want to suppress:
- Fabricating information when uncertain
- Including inappropriate content
- Deviating from specified formats
- Over-explaining when brevity is needed
Negative instructions explicitly block these tendencies. For example:
Without negative instruction: "Summarize this article." → Model might add commentary or tangential details
With negative instruction: "Summarize this article. Do not include your own opinions or analysis." → Model stays focused on factual summary
Strategic Use
Negative instructions work best when:
- Paired with positive instructions (what to do AND what not to do)
- Specific rather than vague ("Do not use technical jargon" vs "Do not be complicated")
- Addressing known failure modes from testing
- Limited in number (too many negatives create confusion)
Example combination: "Extract key dates from the text. Do not infer dates that aren't explicitly mentioned. Do not include approximate timeframes."
Key Takeaway: Negative instructions strategically prevent unwanted model behaviors by explicitly defining boundaries, working best when combined with clear positive guidance and tailored to specific failure modes you've observed.