Hi Daewon Shin , Great question, prompt engineering is key to getting the best results from the OpenAI plug-in in Trados Studio! We already share some examples to help you get started from the plugin, but I include some more here, as well as address your language-dependence question.

Best Practices for Writing Effective Prompts

Be Specific & Direct:
Clearly describe what you want the AI to do. For example:
"Translate the following text from English to German, maintaining the style and terminology of the original text."

Request Style Consistency:
If you want the translation to mimic a particular style or tone (e.g., formal, informal, technical, etc.), mention it.
"Translate the text using a formal and professional tone."

Leverage System Instructions & Example Output:
By default, the system instructions already enforces structure and correct handling of XML tags and locked content, but you can provide your own prompts for extra clarity or context, such as:
"When translating, do not alter any XML tags and ensure terminology from the provided list is used exactly as given."

Give Examples:
Sometimes showing a small example in your prompt helps the AI understand the style or format you want.

Example Prompts

Simple translation prompt:

"Translate the following source text from English to French. Use the provided terminology strictly and maintain the original style as much as possible."

Prompt with formatting/style emphasis:

"Translate this XML-formatted source from German to English. Preserve all tags and formatting precisely. Use the provided terms exactly, and write in a concise, technical style."

About Language-Dependence of Prompts:

OpenAI models perform best with prompts written in English, but they can also understand instructions in other major languages. However: For best results:

• Write your prompt in English even if your source/target text is in another language.
• If you need to, you can write prompts in the source or target language, but results can vary depending on the model and language pair.
• System-level instructions (like tag handling, JSON format) are most reliably understood in English.
  

Additional Tips:

If you want the AI to work similarly to Translation Memory (TM), then you can enable the option 'include target segment with the translation request' if you are explicitly making reference to the existing TM/MT translation of the segment in the prompt and therefore should be included in the translation request.  Then simply include the additional instruction:

"Keep the translation consistent with the existing 'Translation"

let me know how it goes

Hi  Patrick Andrew Hartnett ,

Thank you for the detailed explanation.

I applied the prompt based on your feedback, and the translation results have improved significantly.
However, our team still notices a few issues, particularly with the formatting and symbols from the source text. The output often omits bold text, the '※' symbol, and similar elements.
For example, “※ 참고:” is translated as “Note:”, even though the prompt explicitly instructs the engine to retain the symbol and formatting.

Additionally, when a menu or function name appears in bold in the source text, the engine tends to generate awkward or incorrect translations.
For instance, given a source sentence such as:
“This is an example 기능을 사용하면, ~할 수 있습니다.”
(which means “You can do ~ if you use the This is an example function.”)
the AI sometimes outputs “This is an example” or “This is an example. You can do ~”, which are clearly incorrect and do not make any sense.

Could you please advise how we can further reduce these errors, which seem to be caused by engine errors/limitations?

Thank you

Hi Daewon Shin , thank you for this feedback. 

The quality of the result also depends on the model you are using to interpret yr prompt and generate the translations. Which version of Trados are you using and Plugin?

Can you provide me with the logs from the plugin (not Trados)?

I'll add this to our board so that we can track it with the team.

Note: I see you have already provided me logs via e-mail.  If this content is included in those logs, then no need to resend. 

 Hi Patrick Andrew Hartnett ,

yes, I sent 3 log files, but I am not sure if those files contain this issue. I can send you some more with issues generated on the controlled environment.
Question: plugin logs are different from the Trados logs that I sent you before?

And about Trados version, I have been testing OpenAI on Trados in both 2022 and 2024, and also with peer translators for different languages, but issues (pre-translation, termbase, etc.) mostly occur in Trados 2022. However, this prompt issues are quite identical in both 2022 and 2024. 

Thank you for your help. I hope that I can optimize the settings with your help to get the most out of it.

Thank you very much. I will look forward to your feedback.