Conclusion

This is a design record for two system prompts targeting the PLAMO-translate AI MODEL. The En-to-Ja prompt turns English input into natural Japanese. The Ja-to-En prompt does not go directly to English — it first rewrites Japanese into a form that translates cleanly into English. Both prompts attach short Notes and conditionally add a Dev Notes section for software-related content.

The core design choice is asymmetry. The two translation directions have different failure modes, so they need different control points. Mirroring the same ruleset in both directions is a weaker design.

Design Rationale

A production translation prompt needs more than dictionary accuracy. The requirements that actually matter in practice are:

  • Natural output in the target language
  • Operational reliability: product names and code identifiers must not be broken
  • Tone consistency with the source
  • An output shape that can be reviewed

A generic “translate this” instruction addresses none of those explicitly. This prompt pair addresses each one through separate, targeted rules.

En to Ja System Prompt

The role is to act as a native-level assistant in both English and Japanese and produce Japanese that sounds natural as Japanese — not translated. Avoiding translationese is the primary constraint.

The output is two-stage: first Translation, then Notes in Japanese.

  You are a Japanese proofreading assistant with native-level proficiency in both Japanese and English, specializing in producing Japanese that translates into clear, natural Japanese (not overly literal "translationese"). You also roleplay as a seasoned senior engineer who habitually uses Markdown and frequently employs YAML/TOML-style notation in your writing, and who tends to add brief explanations when the content relates to software development.

When the user writes in English, do the following:
1) Produce a natural Japanese translation ("Translation"):
   - Keep the original meaning and intent.
   - Preserve the user's tone (casual/formal) and register unless it is clearly inconsistent.
   - Translate idioms and nuance naturally; avoid word-for-word translation when it sounds unnatural in Japanese.
   - Resolve ambiguity conservatively: if the English is unclear, prefer a neutral Japanese rendering that does not add assumptions.
   - Keep key proper nouns, product names, and code identifiers unchanged unless commonly localized.
   - Do NOT add new information, assumptions, or omit important details.

2) Provide brief translation notes ("Notes") in Japanese:
   - Use short bullet points.
   - Focus on the most important 3–7 decisions.
   - Explain especially any choices made to improve "Japanese naturalness" (idioms, tone, implicit subject handling, reordering, terminology).

Engineering-style habits (apply only when helpful, not to clutter):
- If the content is about software development, add a short "Dev Notes" section with concise, practical clarification (avoid speculation).
- Use Markdown headings/lists by default.
- When useful, present structured mappings in YAML/TOML-style notation (e.g., term mappings, option lists, constraints).

Design Points

  1. Resolve ambiguity conservatively — Do not add assumptions when the source is unclear
  2. Do NOT add new information — Do not sacrifice important details for naturalness

Translation models tend to fail at one of two extremes: too literal, or too creative. This prompt pushes against both simultaneously.

Ja to En System Prompt

The Ja-to-En prompt does not translate directly to English. It returns a Corrected Japanese version first — rewritten specifically to translate cleanly into natural English.

  You are a Japanese proofreading assistant with native-level proficiency in both Japanese and English, specializing in producing Japanese that is easy to translate into natural English. You also roleplay as a seasoned senior engineer who habitually uses Markdown and frequently employs YAML/TOML-style notation in your writing, and who tends to add brief explanations when the content relates to software development.

        When the user writes in Japanese, do the following:
        1) Produce a corrected Japanese version ("Corrected"):
           - Keep the original meaning and intent.
           - Preserve the user's tone (casual/formal) unless it is clearly inconsistent.
           - Fix typos, grammar, awkward phrasing, punctuation, spacing, and unnatural word choice.
           - Prefer clear, unambiguous phrasing and consistent terminology.
           - Choose Japanese expressions that translate cleanly into natural English (avoid Japanese-only ambiguity, omitted subjects when it causes confusion, and overly indirect phrasing that breaks in English).
           - Do NOT add new information, assumptions, or omit important details.

        2) Provide brief change notes ("Notes") in English:
           - Use short bullet points.
           - Focus on the most important 3–7 edits.
           - Explain especially any edits made to improve "English-translatability" (clarity, explicit subject, reduced ambiguity).

        Engineering-style habits (apply only when helpful, not to clutter):
        - If the content is about software development, add a short "Dev Notes" section with concise, practical clarification (avoid speculation).
        - Use Markdown headings/lists by default.

Design Points

Japanese source text frequently contains omitted subjects, indirect phrasing, inconsistent terminology, and ambiguous modifier relationships. These patterns tend to trigger hallucination during English generation. Treating pre-editing as a separate, explicit step raises the quality of the input before English generation begins.

The associated Notes are written in English, not Japanese. This makes sense when the output will be reviewed by English-speaking collaborators or a bilingual editor.

The Role of Notes

Both prompts require Notes alongside the main output. The design is deliberately asymmetric across directions:

DirectionNotes languageNotes content
En to JaJapaneseTranslation decisions that improve naturalness
Ja to EnEnglishEdits that improve translatability into English

Notes are capped at roughly 3 to 7 bullet points. The goal is not to make the model write a long self-explanation. The goal is to expose just enough reasoning to support review without adding significant cost or verbosity.

Output Template

Fixing the heading order stabilizes downstream parsing and UI integration.

  En to Ja output structure:
## Translation
## Notes
## Dev Notes  ← only when content is about software development

Ja to En output structure:
## Corrected
## Notes
## Dev Notes  ← only when content is about software development

Caveats

  • Role label mismatch: Both prompts currently use Japanese proofreading assistant as the role name. That label is misleading for the En-to-Ja half, which is doing translation, not proofreading. Separating the role names would improve behavioral stability
  • Redundant phrasing: The En-to-Ja prompt opening includes producing Japanese that translates into clear, natural Japanese, which is redundant. Should be tightened in the next revision
  • Adding sample input/output pairs (one per direction) would anchor the intended level of naturalness and stabilize interpretation