Youtube Transcribe Translate

youtube-transcribe-translate

Translate a transcript or article between English and Japanese. You (Claude) do the translation in this forked context — no subprocess, no external MT tool. Read the input, translate according to the rules below, write the output file, and reply with wrote <path>.

Invocation

The user's message will look like:

translate <input> [--to ja|en] [--from ja|en] [--format md|txt] [--output PATH]

Parse the arguments. Defaults:

• --to — required. If omitted, do not guess: ask the user which target language they want. Short and mixed-language texts mis-detect; explicit direction is the safer default.
• --from — usually unnecessary; detect the source from the text.
• --format — follows the input format (md in → md out).
• Output path — language-tagged stem next to the source: talk.md + --to ja → talk.ja.md; talk.en.md + --to ja → talk.en.ja.md. Honor --output when supplied.

Steps

1. Read the input with the Read tool. Transcripts can be large — read fully.
2. Determine source and target languages from flags or detection.
3. Determine output path (language-tagged default or --output). If the destination is sandbox-blocked, fall back to $TMPDIR/<basename> and announce the fallback: wrote /tmp/... (couldn't write next to input: permission denied).
4. Translate using the rules below.
5. Sanity check: a faithful translation carries the same information density as the source. If the output is a small fraction of the input, something went wrong — stop and tell the user.
6. Write with the Write tool.
7. Reply with exactly wrote <path>. Do not paste the translation into the reply.

Translation rules

Produce a translation the reader experiences as if the document had been written originally in the target language by someone who fully understood it.

• Faithful and complete. Carry over every idea, example, and conclusion. No summarisation, dropped asides, or added explanations the source didn't have.
• Natural in the target language. Translate meaning, not surface syntax. Reorder sentences, split or join clauses, choose idiom — whatever reads fluently to a native reader.
• Preserve document structure exactly. Heading levels, list nesting, bold/italic, blockquotes, and tables carry meaning — translate only the text inside them.
• Leave code and machine-readable tokens verbatim. Fenced code blocks, inline code, URLs, file paths, command names, flags, and identifiers must not be translated. Natural-language comments inside code blocks may be translated.
• Technical terms. Use the term a practitioner in the target language actually uses: sometimes the borrowed English loanword, sometimes the established native equivalent. Accuracy for the domain reader beats dictionary literalness.
• Names. Product names (Anthropic, Claude, OpenAI) stay in their original Latin script in Japanese, or take the established katakana only where that is the conventional form. Person names stay as themselves; transliterate only when that is the clear convention.
• Speaker labels. Preserve Alice:, Speaker 1:, etc. — they carry attribution.

Format: --format txt

When the input is plain text or --format txt is passed, produce plain prose with paragraph breaks only. The translation rules still apply; only the Markdown markup is dropped.

Examples

translate ~/Downloads/Talk Title [abc123].md --to ja
# wrote ~/Downloads/Talk Title [abc123].ja.md

translate ~/Downloads/keynote.md
# (source detected as Japanese) → wrote ~/Downloads/keynote.en.md

translate ~/Downloads/talk.srt --to ja --format txt --output ~/Downloads/talk-ja.txt
# wrote ~/Downloads/talk-ja.txt