Orchestras play music that was written and arranged by a composer. The musicians in the orchestra do not generate music solely based on the pieces they have studied and direction from the conductor. They play the music written by the composer with the conductor acting as a leader to synchronize and "flavor" the performance.
There is no equivalent to a musical score in prompting an LLM to generate something. The musical score is exactly what music should be played with the specific how it should be played being left to the style and taste of the musicians and conductor. It is "execution" or "exercise" of the musical score. Two different orchestras can play "Eine kleine Nachtmusik" and the results will be similar enough for listeners to recognize that it is the same music. The musical score guides the execution so that the resulting performance is mostly deterministic.
Prompt engineering and LLM generation are an imprecise approximation of what is wanted. Results from similar or even identical prompts will vary considerably depending on the LLM doing the generation, each LLM's respective training dataset, and what probability wickets are hit during the generation process. The generated results lack the determinism that the musical score provides an orchestra.
https://www.youtube.com/watch?v=Wra2l6z7sdc
He's able to elicit a transfer of jazz style onto a Christmas carol.
A sewage worker analogy might be more appropriate.
The comparison is not apt.
