| name | wjs-localizing-video |
| description | Thin orchestrator for the end-to-end video localization pipeline. Routes to the four focused sub-skills — /wjs-transcribing-audio, /wjs-translating-subtitles, /wjs-dubbing-video, /wjs-burning-subtitles. Use when the user asks for full localization in one go ("帮我把这个西班牙语视频做成中文字幕+配音", "translate and dub this video", "做完整的本地化"). For any individual step (just transcribe, just translate, just dub, just burn), invoke the sub-skill directly — it's faster and the boundary is cleaner. |
wjs-localizing-video (orchestrator)
This skill was split. What used to be one 1300-line catch-all is now 4 focused sub-skills. Use this orchestrator only when the user asks for the full pipeline in one request. For any individual step, route directly to the sub-skill.
The four sub-skills
| Step | Skill | IN | OUT |
|---|
| 1. ASR | /wjs-transcribing-audio | audio / video + source language | source-language SRT |
| 2. Translate | /wjs-translating-subtitles | source SRT + target language | target SRT (punct-bounded) |
| 3. Dub (optional) | /wjs-dubbing-video | video + target SRT + voice | *_<lang>_dub.mp4 (TTS audio swapped in) |
| 4. Burn / mix (optional) | /wjs-burning-subtitles | video + SRT + (optional dub) | final MP4 (one encode) |
Steps 3 and 4 are optional and independent — most "subtitle-only" jobs stop after step 2 + soft-mux via /wjs-burning-subtitles. Most "dub-only" jobs stop after step 3.
When to invoke this orchestrator vs a sub-skill directly
Invoke this skill (wjs-localizing-video) when the user asks for the whole chain in one go:
- "帮我把这个西班牙语视频做成中文字幕和中文配音"
- "translate this Spanish video to English with dubbing"
- "全套：转写、翻译、烧字幕、配音、混音"
Invoke a sub-skill directly when the user asks for one step:
- "转写一下这个音频" →
/wjs-transcribing-audio
- "把这个 SRT 翻译成中文" →
/wjs-translating-subtitles
- "给这个视频配个中文音" →
/wjs-dubbing-video
- "把字幕烧进视频" →
/wjs-burning-subtitles
The sub-skills are more focused and document their input/output contracts tighter. Don't route through this orchestrator when a single step is the whole ask — you'll just be reading docs you don't need.
Progress checklist (do this FIRST, before any sub-skill)
The pipeline has up to four steps and runs for several minutes. The user wants to see live progress, not a wall of silence followed by a finished file. Before invoking the first sub-skill, lay out the planned steps as a TaskCreate checklist. As you start each step, update its task to in_progress; when it finishes, mark completed. Claude Code renders this as a checklist in the UI that ticks off in real time.
Which tasks to create
Decide from the user's ask which subset of the 4 steps will run:
| User asked for | Tasks to create |
|---|
| Full localization (subs + dub + final mix) | ① 转写 ② 翻译 ③ 配音 ④ 烧字幕 + 混音 |
| Subtitles only (no voice change) | ① 转写 ② 翻译 ③ 烧字幕 (or soft-mux) |
| Dub only (no burn) | ① 转写 ② 翻译 ③ 配音 |
| User already has source SRT | skip ①; create only the remaining steps |
| User already has target SRT | skip ① and ②; create only the remaining |
Don't create tasks for steps you won't run — an unchecked item at the end reads as "we forgot," not "we skipped."
How to phrase the task subjects
Use the same labels you'll use in the "Final response template" below, so the in-flight checklist and the completion summary read as the same artifact:
转写 <source-lang> (e.g., 转写 西班牙语)翻译 → <target-lang> (e.g., 翻译 → 中文)配音 <target-lang> (e.g., 配音 中文 (高冷御姐 -8%))烧字幕 + 原声底层 <bed> (e.g., 烧字幕 + 原声底层 0.18)
For activeForm, use the present-continuous variant (转写中…, 翻译中…, 配音中…, 合成中…).
State discipline
- Mark
in_progress only the task currently running (one at a time).
- Mark
completed immediately when the sub-skill returns its output file — don't batch.
- If a step surfaces a clarifying question or fails, leave it as
in_progress so the user can see exactly where the pipeline stopped. Don't mark a half-finished step as completed.
- If the user redirects mid-pipeline (e.g., "skip the dub"), update the remaining tasks: delete the ones no longer applicable, keep the rest. Don't silently drop them.
Canonical end-to-end pipeline (Spanish → Chinese, with dub + bed + burn)
This is the original "validation scenario" — Spanish yoga/spiritual content into Chinese for 微信视频号 / 小红书. Walks through all 4 sub-skills.
Step 1 — Transcribe the Spanish source
Invoke /wjs-transcribing-audio with the video and `--language es`.
The sub-skill handles: chunking, word-level timestamps, cue assembly at punctuation boundaries, loop guard, retry. Output: entrevista.srt (Spanish, source-language).
Step 2 — Translate to Simplified Chinese
Invoke /wjs-translating-subtitles with the Spanish SRT and target `zh-CN`.
The sub-skill handles: re-segmenting cues at punctuation, minimizing filler demonstratives, capping line length, preserving speaker tone. Output: entrevista.zh-CN.srt.
Step 3 — (Optional) Dub into Chinese
Invoke /wjs-dubbing-video with the video + entrevista.zh-CN.srt + a voice ID.
Default for mature contemplative female: zh_female_gaolengyujie_moon_bigtts (Volcano 高冷御姐, --rate -8% +0Hz). If no Volcano credentials: zh-CN-XiaoxiaoNeural --rate -8% --pitch -10Hz. The sub-skill samples first, then commits.
Output: entrevista_zh_dub.mp4.
Step 4 — Burn subs + mix dub over original-as-bed
Invoke /wjs-burning-subtitles with --video entrevista.mp4 --srt entrevista.zh-CN.srt --dub entrevista_zh_dub.mp4
The sub-skill handles: libass availability check, evermeet static-build fallback, Fontsize calibration, frame-check before full render, audio bed mix at 0.18.
Output: entrevista_zh_final.mp4 (ship-ready).
Defaults the user has standardized
These apply across the pipeline; each sub-skill enforces its own slice:
- Chinese ASR routes to 豆包 (Volcano) first; Whisper is fallback. Enforced in
/wjs-transcribing-audio.
- Single voice for the whole dub unless the user explicitly mentions multiple speakers. Enforced in
/wjs-dubbing-video.
- Original audio kept as a low-volume bed (0.15–0.25) under any dub. Enforced in
/wjs-burning-subtitles.
- Burn-in for 微信视频号 / 抖音; soft-mux for everything else by default. Enforced in
/wjs-burning-subtitles.
- Channel CTA name (if added to description): 王建硕 — never a guest's name. (Global rule from
~/.claude/CLAUDE.md; applies when this pipeline is followed by /wjs-uploading-video.)
File naming convention (preserved from the original skill)
input: entrevista.mp4
Chinese pipeline:
source SRT entrevista.srt
Chinese SRT entrevista.zh-CN.srt
Chinese dub (audio only) entrevista_zh_dub.mp4
Chinese final (subs+dub) entrevista_zh_final.mp4
English pipeline:
English SRT entrevista.en.srt
English dub entrevista_en_dub.mp4
English final entrevista_en_final.mp4
Bilingual subtitles:
Spanish + Chinese entrevista.es-zh.srt
Spanish + English entrevista.es-en.srt
three-language entrevista.es-zh-en.srt
BCP-47-style suffixes keep multiple target-language outputs side-by-side and make the target obvious at a glance.
What this orchestrator does NOT do
- It does not re-implement anything. Every step delegates to a sub-skill.
- It does not bundle the scripts.
dub.py lives in /wjs-dubbing-video/scripts/, render.py lives in /wjs-burning-subtitles/scripts/, visual_diarize.py lives in /wjs-dubbing-video/scripts/. If you see stale copies under wjs-localizing-video/scripts/, prefer the canonical sub-skill locations.
Final response template
When the full pipeline completes, respond briefly in the user's language. Match the original "Done" template the user is used to:
已完成：
- 西班牙语转写 (/wjs-transcribing-audio)
- 中文翻译 (/wjs-translating-subtitles)
- 中文配音 (/wjs-dubbing-video, voice: 高冷御姐 -8%)
- 烧入字幕 + 原声底层 0.18 (/wjs-burning-subtitles)
输出：
- entrevista.srt
- entrevista.zh-CN.srt
- entrevista_zh_dub.mp4
- entrevista_zh_final.mp4
不确定片段：
- 00:01:23–00:01:26 背景噪音较大，原文可能不完全准确。
If there are no uncertain parts, drop the second list.
See also
/wjs-segmenting-video — cut long-form video into stand-alone short clips (uses its own SRT slicer; orthogonal to this pipeline)./wjs-overlaying-video — HTML/CSS captions on a clip via HyperFrames. Don't combine with /wjs-burning-subtitles — pick one caption system per output./wjs-uploading-video — push the final MP4 to YouTube./lark-minutes — alternate Chinese-only transcript path via 飞书妙记 when local 豆包 ASR isn't wired up.
