Prompt Transition

End a prompt by firmly pivoting from explaining the problem to solving it, so the model completes an answer instead of adding more framing.

The final part of a prompt should firmly pivot from explaining the problem to solving it — this is the part the model is meant to complete, and it should not just add more (possibly fabricated) context of its own.

In chat interfaces this can be as simple as ending on a question mark; RLHF training has taught such models to answer the last stated, or even implied, question. Some commercial chat platforms auto-signal via the API when the assistant should begin responding, removing the need to engineer this explicitly.

Completion models need more explicit guidance, since nothing in the API tells them a turn has ended. The most common technique, sometimes called inception, is to shift from problem-poser to problem-solver directly in the prompt text: begin writing the answer yourself and let the model continue it as if it had authored the opening — even a trailing opening quotation mark can serve this purpose. Beyond driving the transition itself, inception improves instruction compliance, eases downstream parsing, and removes ambiguity about how the response should open. The refocus / sandwich technique and the transition can often be merged: write the beginning of the answer as a restatement or summary of the problem, and let the model supply the actual answer from that point. This pairs with templated prompt tasks, which use prefix/suffix framing to force extractable completions.

---

Sources

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