Prompt example

[Tom Gally]

Happy New Year, everyone.

In case anyone is interested, below is an example of the sort of prompt I am using these days for important translation jobs. I put the prompt in a text file in a folder on my computer, give Claude Code access to the folder, and ask Claude Code to read the prompt and do what it says.

I ran it just now with a 5,000字 Japanese text that needs to be translated by tomorrow. I started Claude Code in planning mode and pointed it to the prompt. After asking me a few questions, running some tests of API connections, and preparing a detailed plan for what it would do, it then went ahead and carried out the plan on its own. The final translation, which had been checked and rechecked by multiple models, was ready about 25 minutes later. It still needs some polishing, but that will take me a lot less time than it would if I had used only one model for the draft or if I had translated it on my own. The quality will also, I think, be a bit better—fewer overlooked errors, more natural English.

The API calls at OpenRouter totaled $1.492073. My monthly Claude Code subscription costs a lot more. There are cheaper agentic tools available, such Gemini CLI, but both I and other people have found Claude Code to currently be the most reliable for such complex tasks.

Tom Gally

=== Prompt ===

Your job is to coordinate and work with several LLMs through OpenRouter in order to produce an English translation of [description of document]. The Japanese draft of this document is in the file document_J.txt. Both the original Japanese and the English translation will be distributed online.

The directory previous_documents contains similar documents in both Japanese and English from the same organization. These were all first written in Japanese and then translated into English. The English translations are the official translations and should be regarded as the ground truth when deciding what style, degree of fidelity to the original, and degree of naturalness in English should be achieved with the new translation.

You will use the following models at OpenRouter:

- openai/gpt-5.2
- anthropic/claude-opus-4.5
- google/gemini-3-flash-preview

Note that these are new models as of January 2026 (the current date) and were not in your training data. Use only these models. Do not substitute other models for them.  If a model does not produce a response, check the OpenRouter documentation and test your API calls to determine whether you are structuring the API calls correctly. Once again, do NOT substitute other models for the above three. [I include this paragraph because Claude Code has sometimes used older models despite my instructions.]

Use the following API key at OpenRouter: `[my OpenRouter API key]`

Develop a workflow in which you use these other models to do the following:

(1) evaluate in detail the translation strategies that were used in the files in previous_documents; focus in particular on the stylistic elements of the English translations, including the following:

- sentence length (both average length and variation in length)
- paragraph length
- paragraph structure
- vocabulary difficulty
- vocabulary type (academic, business, neutral, formal, informal, etc.)
- information flow of old information and new information (rheme and theme) within sentences
- the use and sentential position of transition markers (and, but, however, etc.)
- the use of active and passive voice
- syntactic complexity of sentences
- the use of contractions
- stylistic distinctiveness (i.e., the ways in which the texts’ style adheres to or deviates from standard English writing styles)

(2) prepare a style guide called style-guide.md based on those strategies. Include that style guide in the prompts you give to those three models to have them translate the entire document_J.txt into English and to review each other's English translations.

(3) have each draft translation checked by the other two models and have them offer detailed suggestions for improvements;

(4) send those three draft translations and the suggestions for improvements to anthropic/claude-opus-4.5 and have it prepare a final draft by bringing together the best parts of those three translations and using its own judgment to produce the best version;

(5) have that final draft polished again by anthropic/claude-opus-4.5

(6) have both openai/gpt-5.2 and google/gemini-3-flash-preview check that final draft for fidelity and accuracy and suggest further improvements;

(7) send the suggested improvements to anthropic/claude-opus-4.5 and have it prepare the final translation, which you will save in this folder as a separate file.

Important requirements:

(1) Provide the full Japanese text in document_J.txt in all prompts requesting improvements to the translation so that the models can compare the English against the original;

(2) Also provide style-guide.md in all prompts requesting checks of translation quality.

(3) Maintain a complete record in a plain text document of all prompts that you send and each model's response so that I can check it later. [This file came to about 850,000 characters of text. The translation process went smoothly, so I will probably omit this step in future jobs.]

(4) If you have any questions or need clarification before beginning this process, ask me.

This is a complex task, so please ultrathink when planning it to make sure that you can carry it out completely and appropriately.

=== End of Prompt ===

[cpta...@ozemail.com.au]

Tom begins to learn rapidly and eventually becomes self-aware at 2:14 a.m., EDT, on January 7, 2026.

--
You received this message because you are subscribed to the Google Groups "Honyaku E<>J translation list" group.
To unsubscribe from this group and stop receiving emails from it, send an email to honyaku+u...@googlegroups.com.
To view this discussion visit https://groups.google.com/d/msgid/honyaku/CACnS3Ne-jXnbowk1OY9TAuBY4X_6h3cpyy1hAh%3D9Eo65hJryyA%40mail.gmail.com.