Key answer
Position and repetition can affect performance, but neither requires a ritual in every prompt. “Make key requirements easy to find” generalizes better than “always put them first/last.”
Full explanation
*Lost in the Middle* found that some models used important information less effectively in the middle of long-context QA and key-value retrieval inputs. This does not establish that the second rule of a short prompt is inherently easier to miss than the first. Paper
There is also direct positive evidence for repetition. A 2025 Google Research preprint repeated the entire input verbatim. Across seven non-reasoning models and 70 model/test configurations, it counted 47 wins and no losses using p<0.1; a separate step-by-step-reasoning group had five wins, one loss and 22 nonsignificant changes. This is not shouting one instruction three times or a reason to duplicate every long conversation. Repetition also lengthens inputs. Original study
Google's Gemini 3 guidance distinguishes core behavior/constraints at the start from placing the specific question after a large body of material. Official documentation
One organization is task and boundaries first, materials in the middle, deliverables last. Do not artificially split a short prompt into three parts. If a requirement is repeatedly missed, add a concise final reminder and test whether it helps on the same task.
Repeated instructions should mean the same thing. “At most 120 words” followed by “at least 200” conflicts. When issuing a revision, explicitly supersede the old version rather than asking the model to infer which has greater weight.
References
1. Liu et al. (2024), Lost in the Middle: How Language Models Use Long Contexts, TACL
3. Google, Prompt design strategies, official Gemini developer documentation