Prompt repetition copies the full prompt and places the copy before the model responds. This gives the model a second pass over the same context without changing the requested output format. section 4.5
Use repetition when:
- First-response accuracy matters in a nonreasoning configuration, especially when relevant details appear early or midway through a long prompt. section 4.5.1
- Skip it when reasoning is enabled, the prompt is near the context-window limit, or token cost matters in a high-volume workflow. Repetition doubles input tokens and may increase latency. section 4.5.1
Choose a logging approach
A prompt describes four logging strategies for a high-throughput microservice, then asks which has the lowest performance overhead. With repetition, the complete prompt—including the strategy descriptions and question—is included twice. The example selects asynchronous logging and explains that queuing moves writes off the request path; it also notes possible log loss, ordering ambiguity, and backpressure trade-offs. section 4.5.2
| Variant |
Structure |
Best used for |
|---|---|---|
| Standard |
Prompt repeated twice, back-to-back |
Most tasks; clean and unobtrusive |
| Verbose |
Prompt, then “Let me repeat that:”, then prompt again |
When explicit signposting helps the model register the repetition |
| x3 |
Three full repetitions with transition phrases |
Long list retrieval or tasks with extreme middle-context sensitivity |
Variants range from two back-to-back copies to signposted repetition or three copies for tasks especially sensitive to details in the prompt’s middle. section 4.5.2
