Preserving Tone and Formatting in AI Translation

English original: "Our coffee goes stale fast — that's the whole
  point. Freshness this real means no shortcuts."

Overly literal translation: technically correct word-for-word, but
  reads stilted and unnatural in the target real language

def translate(text: str, target_language: str, tone: str = "warm and direct") -> str:
    client = get_client()
    response = client.messages.create(
        model="claude-sonnet-5",
        max_tokens=1024,
        system=(
            f"Translate the following text into {target_language}. "
            f"Prioritize natural, idiomatic phrasing over literal, "
            f"word-for-word translation. Maintain a {tone} tone, "
            f"consistent with Bright Leaf Coffee's real brand voice. "
            f"Preserve product names exactly as written. Output ONLY "
            f"the translation."
        ),
        messages=[{"role": "user", "content": text}],
    )
    return response.content[0].text

markdown_description = """
## Ethiopian Light Roast

**Tasting notes:** blueberry, floral

- Origin: Yirgacheffe, Ethiopia
- Roast level: Light
"""

def translate_preserving_markdown(text: str, target_language: str) -> str:
    system = (
        f"Translate the following Markdown text into {target_language}. "
        f"Preserve ALL Markdown syntax exactly — headings, bold, "
        f"bullet points — translating only the actual real content, "
        f"never the formatting characters themselves."
    )
    # ...same real request structure as before

def verify_translation_quality(original: str, translated: str, target_language: str) -> str:
    client = get_client()
    response = client.messages.create(
        model="claude-sonnet-5",
        max_tokens=512,
        system=(
            "You are reviewing a translation for quality. Translate "
            "the provided translated text back into English, then "
            "note any real, meaningful differences in meaning from "
            "the original — ignore minor, expected phrasing "
            "differences."
        ),
        messages=[
            {
                "role": "user",
                "content": f"Original English: {original}\n\nTranslated back from {target_language}: {translated}",
            }
        ],
    )
    return response.content[0].text

Automated back-translation: catches OBVIOUS, real meaning shifts
Does NOT catch: subtle cultural tone mismatches, idioms that
  technically translate correctly but land oddly for a real,
  specific target audience