Structuring Policy Prompts

Structuring Policy Prompts

Policy prompts should have four separate sections.

1. **Instruction:** what the model MUST do and how the model should answer.
2. **Definitions:** concise explanations of key terms.
3. **Criteria:** distinctions between violating and non-violating content.
4. **Examples:** short, concrete instances near the decision boundary. It’s important to have both examples of what you want to classify, and what you do not want to classify

Because oss-safeguard is tuned for structured moderation, it expects explicit instructions for how to respond. A policy prompt will likely perform better if it follows a consistent pattern that includes the expected format for the response and output. The harmony format’s structured channels allow oss-safeguard to reason through these sections before emitting only the final label:

```markdown