End-to-end video translation pipeline: transcribe → translate → refine subtitles → embed.

/translate-video /path/to/video.mp4 he --regular
/translate-video /path/to/video.mp4 he --shorts

• $1 — video file path (required)
• $2 — target language code (default: he). See references/languages.md
• $3 — --shorts (TikTok/Reels) or --regular (YouTube/tutorials). If omitted, ask the user.

If audio > 25MB, extract first:

ffmpeg -i "$VIDEO" -vn -acodec libmp3lame -ab 128k "$AUDIO.mp3" -y

Transcribe with word-level JSON (always include --json):

cd ~/.claude/skills/transcribe/scripts && [ -d node_modules ] || npm install --silent
npx ts-node transcribe.ts -i "$INPUT" -o "$BASENAME.srt" --json

Produces: {basename}.srt, {basename}.md, {basename}_transcript.json

Read .md for full context. Translate the .srt — preserve all timestamps and index numbers exactly. See translation rules in references/modes.md.

Read references/modes.md for full rules.

--shorts: Fix text only, preserve all timestamps. No merging.

--regular: Merge into full sentences using word-level timestamps.

1. Plan subtitle groups from .md (word counts per group)
2. Fill GROUP_SIZES in scripts/build-timestamps.py and run it
3. Replace English text in output SRT with translated text

Enforces MAX 2 lines, MAX chars/line (38 for --shorts, 42 for --regular):

python3 ~/.claude/skills/translate-video/scripts/postprocess.py "$SRT" 42

python3 ~/.claude/skills/translate-video/scripts/rtl-fix.py "$SRT"

~/.local/bin/ffmpeg-ass -i "$VIDEO" \
  -vf "subtitles=$SRT:force_style='FontSize=24,PrimaryColour=&H00FFFFFF,OutlineColour=&H00000000,Outline=2,Shadow=1,Alignment=2,MarginV=30'" \
  -c:v libx264 -preset fast -crf 23 -c:a copy "$OUTPUT" -y
open "$OUTPUT"

⚠️ Do NOT use Docker ffmpeg for long videos on ARM Mac — x86 emulation is ~100x slower.