(Note) The following commentary was was initially composed in Japanese and then translated into English, on translation from English to Japanese. While the discussion is primarily about translation from English to Japanese, the insights could be applicable to translation between any pair of languages.
The current prompt I use for GPT-4, which seems effective so far, is:
(Start)
"Determine the accuracy of the comprehension of the English text in the following translations, and carefully comment on each sentence considering the overall specified range, whether the vocabulary, grammar, and expressions in the translated Japanese are appropriate. If a better alternative to the presented translation exists (if not, it is unnecessary), also offer it in a form that aligns as closely as possible with the provided Japanese translation (with minimal revisions)."
(End)
In practice, here's what happens: even with the same prompt, the reaction to GPT-4's response can vary based on the user, leading to significantly different results. Despite using the same source text and prompt for a task or research, the final translation can vary based on who's in charge.
Though not identical, it's reminiscent of how translations of the same English text can differ, even when using the same dictionary. It seems that the user's personal style influences the outcome.
As a side note, to manage usage limits, I regularly switch between chatGPT-4 and chatGPT-3. Sometimes, I embarrassingly mix them up, and this often results in much less accurate responses when I accidentally use GPT3.5.
