Ensuring a clear distinction between prompt instructions and input data is critical for maintaining system stability and security, particularly to prevent prompt injection attacks where the model misinterprets data as executable commands. To achieve this, developers effectively create "fencing" around the data using explicit delimiters like triple quotes, hashtags, or XML tags, which serve as visual and logical boundaries that tell the model exactly where the input begins and ends. Furthermore, utilizing structured formats like JSON or leveraging API-level role separation (distinguishing between "System" instructions and "User" content) provides a robust architectural hierarchy, ensuring the model processes the input solely as content to be analyzed rather than instructions to be obeyed.
Techniques for Differentiating Prompts and Data
| Method | Description | Example Implementation |
|---|---|---|
| Delimiters | Using repeating special characters to create a "fence" around the input data. | Summarize the text delimited by triple quotes: """[Input Data]""" |
| XML/HTML Tags | Enclosing data within specific opening and closing tags to define the data's scope. | Analyze the sentiment of the text inside the <review> tags: <review>[Input Data]</review> |
| Role-Based Separation | Using API parameters to separate high-level instructions from user input. | System Role: "You are a translator." User Role: "[Input Data]" |
| Structured Formats | Encapsulating data within a rigid schema like JSON to parse content programmatically. | Extract entities from the following JSON object: {"content": "[Input Data]"} |
| Explicit Headers | Using capitalized labels and line breaks to demarcate sections. | INSTRUCTIONS: Classify the text below. |
Ready to transform your AI into a genius, all for Free?
Create your prompt. Writing it in your voice and style.
Click the Prompt Rocket button.
Receive your Better Prompt in seconds.
Choose your favorite favourite AI model and click to share.