Mark the boundaries
When a prompt mixes instructions with a block of user text, the model can confuse the two. Delimiters are clear markers, such as labeled sections or fenced blocks, that fence off the data so the model knows where each part begins and ends.
What delimiters buy
- Separation of the instruction from the content to act on.
- Clarity about which span is the literal input versus the command.
- Resistance to text inside the data that looks like a new instruction.
A defense against injection
If user supplied text contains an order like ignore the above, an undelimited prompt may obey it. Wrapping that text in a clearly marked block and telling the model to treat everything inside as data to analyze, not instructions to follow, reduces that risk.
Keep markers distinct
Pick delimiters unlikely to appear in the data, label sections clearly, and reference them by name in the instruction. Consistent structure makes long prompts easier for both the model and you to follow.
Key idea
Delimiters fence off data from instructions so the model knows what to treat literally, improving clarity and resisting injected commands hidden inside user supplied text.