Sandwich Technique

State the goal at both the start and the end of a long prompt so a refocus after intervening context restores precise operational intent.

The sandwich technique places what you want the model to do at both ends of the prompt: an introduction that sets document type and focus, then a long middle of context, then a refocus that re-asks the question with precise operational detail (and often output-format constraints). In a chat transcript this can look like a system message plus a first user turn stating the goal, a long back-and-forth of context-gathering turns, then a final assistant turn signaling something like "I believe this is all the information I need," followed by a final user turn re-asking the question.

Introduction vs refocus: the introduction stages ("thinking about book suggestions for X"); the refocus commands ("recommend the best next book, narrative prose, currently available"). After heavy context gathering — chat turns, retrieved blocks, tool traces — the original ask is easy to bury in the Valley of Meh, the weak band lost in the middle attention creates; the trailing restatement recovers it. A short extra refocus at the very end can lock format. Often merge the refocus with the prompt transition by beginning the answer as a problem restatement the model must complete.

---

Sources

• Prompt Engineering for LLMs (Berryman), ch. 6