What is negative prompting and how do you instruct models what not to do?

Short answer: Negative prompting explicitly specifies prohibited behaviors, restricted topics, or disallowed words, preventing model hallucinations, off-topic rambling, and safety policy breaches.

Detail

LLMs process text by predicting tokens that align with prompt semantics. Naive negative instructions like "Do NOT think of a pink elephant" often inadvertently increase attention weight on the banned concept.

Weak Negative Prompt:   "Don't generate conversational intros." (Model often outputs "Sure! Here is...")
Strong Negative Prompt: "Respond strictly with the target JSON object. Zero conversational filler or introductory text."

Best Practices for Negative Directives

1. Positive Replacement: Instead of solely banning a behavior, instruct the model on the positive alternative (e.g. "Instead of bullet points, use a single continuous paragraph").
2. XML Tag Isolation: Wrap negative rules inside distinct structural blocks (<disallowed_actions>).
3. Structured Validation Guardrails: Pair negative prompts with Pydantic output validation to catch edge-case violations.

Example

Python prompt comparison highlighting positive framing of negative constraints:

# Weak
prompt_weak = "Write a summary. Do not include pricing or competitor names."

# Strong
prompt_strong = """Write a product summary.
DISALLOWED:
- Do NOT mention pricing or competitor brand names.

REQUIRED FORMAT:
- Focus exclusively on technical product specifications and core features."""

Interview tips

• Explain why LLMs sometimes ignore negative constraints (due to token-level self-attention activating mentioned concepts).
• Mention combining negative prompts with output guardrail classifiers like Llama Guard or regex validators.

Related Concepts

• [[What is model jailbreaking and how do safety classifiers block it?]] (#184): What is model jailbreaking and how do safety classifiers block it?
• [[What is assertion testing in LLM unit tests?]] (#173): What is assertion testing in LLM unit tests?
• [[How to explain LLM hallucination mitigation in a system design interview?]] (#195): How to explain LLM hallucination mitigation in a system design interview?

⬅ Back to Prompt Engineering · All topics