A delimiter is a boundary marker around embedded content in a prompt. It helps distinguish text to analyze from instructions to follow, especially when a template receives variable input at runtime. section 2.7
Choose and place delimiters deliberately:
- Use triple-backtick fences for ordinary contextual content; XML-style tags are another option. section 2.7
- Choose markers unlikely to occur in the content. If an input already contains triple backticks, use a different boundary. section 2.7.1
- Keep the style consistent within a prompt and make the boundaries between instructions and inputs explicit. section 2.7.3
Incident-report template
# Instructions Summarize the incident report. Cover impact, root cause, mitigation, and next steps. Use a neutral, operational tone. # Incident Report ```text {{incident_report}} ``` # Output Format Impact; Root Cause; Mitigation; Next Steps. The fenced placeholder marks the changing report as content to summarize. The instructions and output format remain outside it. section 2.7.1 section 6.1.2
For several inputs, label and delimit each one separately. A production-bug prompt can fence an error log, a change summary, and relevant code before asking for a likely root cause. Separate blocks make the template easier to populate, maintain, and debug. section 4.3.1
Delimiters reduce instruction-mixing risk, but they do not secure a prompt on their own. For external text, pair them with explicit instructions that treat the enclosed content as data, plus basic input hygiene. Add delimiters when they help solve the task rather than forcing them into every prompt. section 2.7.3 section 2.8
