Skip to content

jianshuo/claude-skills / wjs-localizing-video

wjs-localizing-video

Verified

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.

FileKindSize
.gitignoreother611 B
.publish-manifest.jsonconfig45.1 KB
LICENSEother1.0 KB
README.mdmarkdown9.8 KB
SKILL.mdskill md9.2 KB

.publish-manifest.json

View on GitHub
{
  "skill": "translate-video",
  "repo_url": "https://github.com/jianshuo/translate-video",
  "last_published_at": "2026-05-10T15:04:42Z",
  "last_published_sha": "41eee194e41448a00d885944550fa1663df323a9a2594dbdd373848ec7def132",
  "last_published_md": "---\nname: translate-video\ndescription: End-to-end video localization. Transcribe spoken audio in any language Whisper supports (Spanish, English, Portuguese, French, Italian, Japanese, Korean, etc.), translate into a chosen target language (Simplified Chinese and English are first-class; other targets work via the same pipeline if a TTS voice is available), generate punctuation-bounded SRT subtitles, optionally burn them into the video, and optionally produce a time-aligned voice dub. Defaults to single-speaker — uses one voice for the whole video. Multi-speaker dubbing (different voice per person) is an opt-in advanced mode triggered only when the user explicitly says the source has multiple speakers. TTS routes by voice ID — Volcano (豆包) for Chinese, edge-tts neural for any language. Preserves the original audio as a low-volume bed under the dub when desired. Bundled scripts in `scripts/`: `dub.py` (TTS + time-align), `render.py` (burn + mix + final), and `visual_diarize.py` (opt-in mouth-movement speaker detection). All major behaviors are flag-controlled.\n---\n\n# translate-video\n\n## Purpose\n\nEnd-to-end video localization pipeline. Given a video with spoken\naudio in **any language Whisper recognizes**, this skill produces:\n\n1. A timestamped transcript SRT in the source language.\n2. A translated SRT in the user's chosen target language, segmented\n   at punctuation boundaries (no mid-sentence breaks).\n3. Optional: hard-burned subtitles in the chosen target language.\n4. Optional: a time-aligned TTS voice dub in the target language,\n   with the original audio optionally preserved as a low-volume bed.\n\nOutputs are SRT and MP4 — usable directly in Final Cut Pro,\nPremiere Pro, CapCut, DaVinci Resolve, or `ffmpeg`.\n\n### Source language\n\nPass `--language es` (or `en`, `pt`, `fr`, `it`, `ja`, `ko`, etc.)\nto whisper to lock detection. Auto-detect can mis-route on short or\nheavily accented clips, so always pin the source explicitly when\nknown.\n\n### Target language\n\nThis skill is fully validated for two targets:\n\n- **Simplified Chinese (zh-CN)** — Volcano (豆包) TTS for dub,\n  Chinese-specific subtitle line conventions.\n- **English (en)** — edge-tts multilingual neural voices for dub,\n  English subtitle line conventions.\n\nOther targets (Japanese, Korean, French, etc.) work mechanically via\nthe same pipeline; the bottleneck is finding a good TTS voice — the\nedge-tts catalog covers most major languages, but cap-test before\npromising.\n\nPicking from user phrasing:\n\n- \"翻成中文 / 中文字幕 / 中文配音\" → `zh-CN`.\n- \"translate to English / English subs / English dub\" → `en`.\n- \"bilingual\" → produce both `.zh-CN.srt` and `.en.srt`; for dubs\n  ask which one to render (or render both).\n- Ambiguous → default to whichever the user has historically chosen\n  in the project; otherwise ask once.\n\nThe canonical worked example throughout this doc is **Spanish → Chinese**\nbecause that was the original validation scenario, but every step\napplies to other source/target pairs unchanged.\n\n### Number of speakers — default to one\n\n**Default: assume one speaker.** Use a single voice for the entire\ndub. This is the right answer for monologues, vlogs, recorded talks,\nnarrator-only clips, and the overwhelming majority of videos people\nask about. Don't run diarization, don't tag the SRT with `[A]`/`[B]`,\ndon't bring up multi-speaker complexity.\n\n**Switch to multi-speaker only when the user explicitly says so** —\nphrasings like \"two people\", \"interview\", \"dialogue\", \"conversation\nbetween\", \"separate the speakers\", \"different voice for each\", or a\ndirect request to do diarization. When triggered, follow the\n\"Multi-speaker dubbing (advanced, opt-in)\" section near the end of\nthis doc.\n\nIf you're unsure whether a video is one speaker or many, ship the\nsingle-voice version first. Adding speaker separation later is\ncheap (just regenerate the dub); shipping confused multi-speaker\noutput by default wastes the user's time.\n\n---\n\n## When to Use\n\nUse this skill when the user asks to:\n\n- Transcribe spoken audio from a video in any source language\n- Translate video speech into Chinese or English (or other languages\n  with available TTS)\n- Add subtitles to a video (soft-muxed or hardcoded/burned-in)\n- Generate `.srt` from audio or video\n- Translate an existing subtitle file into another language\n- Create bilingual or trilingual subtitles\n- Fix, polish, or re-time translated subtitles\n- Produce a voice dub of a foreign-language video, optionally with\n  the original audio kept as a low-volume bed underneath\n\n---\n\n## Input Types\n\nThe user may provide:\n\n- A video file: `.mp4`, `.mov`, `.mkv`, `.avi`\n- An audio file: `.mp3`, `.wav`, `.m4a`, `.aac`\n- A subtitle file: `.srt`, `.vtt`, `.ass`\n- A Spanish transcript pasted into the chat\n- A rough subtitle draft that needs translation or repair\n\n---\n\n## Default Workflow\n\n### Step 1: Inspect the Source\n\nDetermine what the user provided:\n\n- If the user provides a video or audio file, transcribe the Spanish speech.\n- If the user provides an existing Spanish subtitle file, preserve its timestamps and translate line by line.\n- If the user provides only transcript text, translate it and create subtitles only if timing information is available.\n- If no timing information exists, do not invent exact timestamps unless the user explicitly asks for approximate timing.\n\n---\n\n### Step 2: Transcribe the Source Audio\n\nWhen transcribing (any source language):\n\n- Preserve the original meaning.\n- Keep sentence boundaries clear.\n- Do not hallucinate unclear words.\n- Mark unclear audio as `[inaudible]` only when necessary.\n- Preserve names, places, brands, dates, and numbers.\n- If multiple speakers are obvious, label them only when useful.\n- Keep the transcript aligned with timestamps where possible.\n\nPreferred working format:\n\n```text\n[00:00:01.200 --> 00:00:04.800] Spanish transcript here.\n[00:00:04.800 --> 00:00:08.500] Spanish transcript here.\n```\n\n#### Transcription tooling\n\nUse `openai-whisper` via `uvx` so nothing pollutes the system:\n\n```bash\n# Extract 16k mono PCM (faster + smaller than feeding mp4 directly)\nffmpeg -i input.mp4 -vn -ac 1 -ar 16000 -c:a pcm_s16le _audio.wav -y\n\n# Transcribe — outputs .srt next to the input\nuvx --from openai-whisper whisper _audio.wav \\\n    --language es --task transcribe \\\n    --model small --output_format srt --output_dir .\n```\n\nNotes:\n\n- Use `--language es` to lock Spanish; auto-detect can mis-route on\n  short or accented clips.\n- `small` is enough for clean studio audio; jump to `medium` if\n  background noise or overlapping speakers are present.\n- Whisper writes timestamps with `.` milliseconds; the file is still a\n  valid SRT (it auto-converts to `,` on save). If you regenerate the\n  SRT yourself, always emit `,` ms.\n- The first run downloads the model (~480MB for `small`); subsequent\n  runs are cached.\n- Delete the intermediate `.wav` after transcription.\n\n---\n\n### Step 3: Translate Spanish into the Target Language\n\nBoth target languages share the same core principles, but the text\nmechanics (line length, punctuation, registers) differ.\n\n#### Shared principles\n\n- Prioritize meaning over literal wording.\n- Use concise subtitle-style language — viewers read at ~3 wps for\n  Chinese, ~3–4 wps for English; lines that exceed that go off-screen\n  before they can be read.\n- Preserve the tone of the speaker. Casual Spanish → casual target;\n  formal Spanish → formal target.\n- Do not over-translate names, brands, cultural references, or\n  technical terms.\n- Keep numbers, dates, names, and places accurate.\n- If a phrase has no exact equivalent, translate the meaning\n  naturally. No literal/word-for-word constructions.\n- Avoid stiff, machine-translated output.\n\n#### Translating into Simplified Chinese (zh-CN)\n\n- Use natural spoken Mandarin for casual speech, formal Mandarin for\n  formal speech.\n- Use Simplified characters only (do NOT use Traditional Hanzi unless\n  the user explicitly asks).\n- Subtitle lines should be roughly **15 Chinese characters** or fewer\n  per line, max 2 lines per cue (3 only when unavoidable for very\n  long cues — see SRT discipline below).\n- Use Chinese punctuation: 「,」「。」「;」「:」「、」「——」.\n  Never mix English commas/periods into Chinese subtitles.\n- **Minimize filler demonstratives 「这」「那」「这个」「那个」\n  「那份」「那种」「那里」「那样」.** Spanish-to-Chinese MT routinely\n  inserts these because Spanish has overt demonstratives (\"eso, esa,\n  ese, aquello\") that Chinese usually drops. Examples:\n  - \"这把我们带入二元世界的载体\" → \"把我们带入二元的载体\"\n  - \"运用那份能量\" → \"运用这股能量\" if needed, or just \"运用能量\"\n  - \"正是在这合一里\" → \"正是在合一中\"\n  - \"像罪人那样翻滚\" → \"像罪人翻滚\" / \"像罪人般翻滚\"\n  - \"那份精微的觉知\" → \"精微的觉知\"\n  Keep them only when they carry real meaning (deixis, contrast, or\n  fixed phrase like spiritual \"我就是那\" / \"tat tvam asi\"). Default\n  is to delete; add back only if the sentence becomes ambiguous.\n\nExamples:\n\n```text\nSpanish: No pasa nada.\nChinese: 没关系。\n\nSpanish: Vamos a ver qué pasa.\nChinese: 我们看看会发生什么。\n\nSpanish: Me parece una locura.\nChinese: 我觉得这太疯狂了。\n\nSpanish: ¿Qué quieres decir?\nChinese: 你是什么意思?\n\nSpanish: La verdad es que no lo esperaba.\nChinese: 说实话,我没想到会这样。\n```\n\n#### Translating into English (en)\n\n- Use natural conversational English. Avoid translationese\n  (\"It is precisely through entering the body…\" → \"It's by entering\n  the body…\").\n- Lines should be roughly **40–42 characters** or fewer (about 7–9\n  words), max 2 lines per cue. Hard cap 50 chars per line.\n- Use ASCII punctuation: `,` `.` `;` `:` `—` (em-dash). Avoid Unicode\n  curly quotes unless the source needs them — keeps `.srt` portable.\n- For contemplative/spiritual content (the typical Spanish-coach\n  source), prefer plain words over Latinate jargon: \"presence\" over\n  \"manifestation,\" \"wholeness\" over \"totality,\" \"wake up\" over\n  \"awaken to consciousness.\"\n\nExamples:\n\n```text\nSpanish: No pasa nada.\nEnglish: It's nothing.\n\nSpanish: Vamos a ver qué pasa.\nEnglish: Let's see what happens.\n\nSpanish: Me parece una locura.\nEnglish: This feels crazy to me.\n\nSpanish: ¿Qué quieres decir?\nEnglish: What do you mean?\n\nSpanish: La verdad es que no lo esperaba.\nEnglish: Honestly, I wasn't expecting this.\n```\n\n---\n\n### Step 3.5: Re-segment cues at punctuation boundaries\n\nWhisper segments by silence/breath, not grammar. The result almost\nalways has cues that **end mid-sentence** (e.g., \"...es una forma de\naterrizar,\" then next cue starts \"el espíritu en el cuerpo...\"). Any\nTTS that processes one cue at a time will then insert an unnatural\npause exactly where the original speaker did not. The fix is mandatory\nbefore dubbing — and improves on-screen reading too.\n\nApply this regardless of target language. The punctuation set\ndiffers:\n\n- Chinese cues must end at \",\" \"。\" \";\" \":\" \"——\" or \"、\".\n- English cues must end at `,` `.` `;` `:` `—` (em-dash) or, in\n  practice for subtitles, occasionally a single dash. Never end an\n  English cue on a comma-less clause break, and never split inside\n  a phrase like \"kind of\" or \"in order to\".\n\nRules:\n\n- **Every cue must end at a real punctuation mark.** Never let a\n  cue end on a noun, verb, conjunction, or article that flows into\n  the next cue.\n- It is fine (and often necessary) to **split** a single Whisper cue\n  into 2–4 shorter cues, with timestamps interpolated by character\n  position within the original cue's duration.\n- It is fine to **merge** the tail of one Whisper cue with the head\n  of the next when they form one clause — the merged cue inherits\n  the start of the first and the end of the second.\n- Target 3–8 seconds per cue. Cues shorter than ~1.5s feel choppy on\n  screen; cues longer than ~10s usually contain a missed punctuation\n  break.\n\nA typical 2–3 minute talk yields roughly 25–40 punct-bounded cues\nfrom 12–18 raw Whisper cues. This is normal — don't try to keep the\noriginal cue count.\n\nWhen TTS dubbing follows: the punctuation-bounded structure means\neach TTS clip is a complete utterance with proper end-intonation, and\nconcatenating clips sounds natural because every join is at a real\npause point.\n\n---\n\n### Step 4: Generate Chinese SRT\n\nGenerate a valid `.srt` subtitle file.\n\nSRT format:\n\n```text\n1\n00:00:01,200 --> 00:00:04,800\n中文字幕内容\n\n2\n00:00:04,800 --> 00:00:08,500\n中文字幕内容\n```\n\nSRT rules:\n\n- Number subtitles sequentially starting from `1`.\n- Use this timestamp format: `HH:MM:SS,mmm`.\n- Use comma milliseconds, not period milliseconds.\n- Do not overlap timestamps.\n- Preserve the original timing unless adjustment is necessary.\n- Each subtitle should usually be 1–2 lines.\n- Keep each Chinese line readable.\n- Prefer no more than about 18–22 Chinese characters per line when possible.\n- If one subtitle is too long, split it into shorter subtitles when timing allows.\n- Do not add commentary inside the subtitle file.\n\n---\n\n## Chinese-Only Subtitle Output\n\nDefault output should be Chinese-only SRT:\n\n```text\n1\n00:00:01,200 --> 00:00:04,800\n没关系。\n\n2\n00:00:04,800 --> 00:00:08,500\n我们看看会发生什么。\n```\n\n---\n\n## Bilingual Subtitle Output\n\nIf the user asks for bilingual subtitles, use Spanish on the first line and Chinese on the second line:\n\n```text\n1\n00:00:01,200 --> 00:00:04,800\nNo pasa nada.\n没关系。\n\n2\n00:00:04,800 --> 00:00:08,500\nVamos a ver qué pasa.\n我们看看会发生什么。\n```\n\nRules for bilingual subtitles:\n\n- Keep Spanish first.\n- Keep Chinese second.\n- Preserve timing.\n- Avoid adding extra explanations unless requested.\n- Keep both lines short enough to read.\n\n---\n\n## Subtitle Quality Rules\n\nBefore final output, verify:\n\n- Subtitle numbers are sequential.\n- Timestamps are valid.\n- Milliseconds use commas.\n- No subtitle time ranges overlap.\n- Chinese translation is natural.\n- Chinese subtitle length is readable.\n- Speaker tone is preserved.\n- Proper nouns are accurate.\n- Unclear audio is marked honestly.\n- No missing subtitle blocks.\n- No invented content.\n\n---\n\n## Handling Unclear Audio\n\nIf audio is unclear:\n\n- Do not guess aggressively.\n- Use `[inaudible]` for completely unintelligible words.\n- Use `[unclear]` only when part of the speech is uncertain.\n- Mention uncertain sections after the SRT output.\n\nExample note:\n\n```text\nUncertain sections:\n- 00:01:23–00:01:26: background noise makes the Spanish partly unclear.\n- 00:03:10–00:03:13: speaker mentions a name that may be misspelled.\n```\n\nDo not put long uncertainty explanations inside the subtitle file unless the user asks.\n\n---\n\n## Output Formats\n\nDepending on the user request, provide one or more of the following:\n\n1. Chinese-only `.srt`\n2. Spanish-Chinese bilingual `.srt`\n3. Chinese transcript without timestamps\n4. Side-by-side Spanish/Chinese table\n5. Soft-muxed `.mp4` (togglable Chinese subtitle track)\n6. Hardcoded burn-in `.mp4` (always-visible Chinese subtitles)\n7. Chinese voice dub `.mp4`, with three audio modes:\n   - dub-only (replaces Spanish audio)\n   - dub + Spanish bed (Chinese 100%, Spanish at 15–25%)\n   - dub + burned-in subtitles + Spanish bed (full localized cut)\n\nDefault output:\n\n- Chinese-only `.srt`\n- A short uncertainty note if needed\n\nIf the user already has the subtitle file and asks for dub or burn-in,\ngo straight to that — don't regenerate the SRT.\n\n---\n\n## Subtitle Output Modes\n\nThere are two ways to attach the SRT to the video. Pick based on what\nthe user is doing with the file.\n\n### Soft-mux (togglable subtitle track)\n\nPlayer apps (QuickTime, VLC, IINA, mobile players) can show/hide.\nWorks with any `ffmpeg` build — does **not** need libass:\n\n```bash\nffmpeg -i input.mp4 -i input.zh-CN.srt \\\n  -map 0:v -map 0:a -map 1:0 \\\n  -c:v copy -c:a copy -c:s mov_text \\\n  -metadata:s:s:0 language=zho -metadata:s:s:0 title=\"中文\" \\\n  output.mp4\n```\n\n### Hardcoded burn-in (always visible)\n\nRequired for WeChat/抖音/朋友圈 etc. where the player will not honor\nembedded subtitle tracks. Needs an `ffmpeg` built with libass.\n\n**Verify libass is available before promising burn-in:**\n\n```bash\nffmpeg -filters 2>&1 | grep -E \"subtitles|^.. ass \"\n```\n\nIf neither `subtitles` nor `ass` shows up, the build lacks libass.\nHomebrew's default `ffmpeg` formula is often stripped (no\n`--enable-libass`, no `--enable-libfreetype`, no `drawtext`). Don't\nwaste time fighting the comma-escaping inside `force_style` — it will\nfail with `No such filter: 'subtitles'` no matter how the shell quotes\nit.\n\n**Fastest fix on macOS — drop in a static build, no system changes:**\n\n```bash\ncurl -fsSL -o /tmp/ff.zip https://evermeet.cx/ffmpeg/getrelease/zip\nunzip -o /tmp/ff.zip -d /tmp/ff_bin >/dev/null\nFF=/tmp/ff_bin/ffmpeg\n$FF -version | grep -oE -- \"--enable-(libass|libfreetype)\"\n```\n\nThen use `$FF` instead of `ffmpeg` for the render. The brew binary is\nfine for everything else (probe, audio extraction, soft-mux).\n\n**Burn-in render with style overrides.**\n\n🛑 **Checkpoint — confirm before full-render.** Burn-in re-encodes the\nentire video (minutes of CPU on a 5-min clip). Before kicking it off:\n\n1. Render only the first 30s with `-t 30` for a fast preview.\n2. Extract a frame from the longest-line cue (see Fontsize calibration\n   below) and Read it.\n3. Show the user the preview frame + the cue text, ask: \"字号/字体/边距\n   OK 吗?OK 才跑全片。\" Wait for explicit confirmation.\n\nSkip the checkpoint only if the user has already approved a full render\nof this exact video at this exact font config in the same conversation.\n\n```bash\n$FF -i input.mp4 \\\n  -vf \"subtitles=input.zh-CN.srt:force_style='Fontname=PingFang SC\\,Fontsize=12\\,PrimaryColour=&H00FFFFFF\\,OutlineColour=&H00000000\\,BorderStyle=1\\,Outline=2\\,Shadow=1\\,MarginL=20\\,MarginR=20\\,MarginV=40'\" \\\n  -c:v libx264 -crf 18 -preset medium -pix_fmt yuv420p \\\n  -c:a copy output.mp4\n```\n\nInside `force_style`, escape every comma as `\\,` (the filter graph\nparser eats the bare comma as a chain separator). All other special\nchars are fine.\n\n#### Fontsize calibration — critical\n\nlibass scales its internal PlayRes up to the actual video resolution.\nThe number you pass is **not pixels** in the output. As a starting\ncalibration on a 544×960 vertical phone video, `Fontsize=22` rendered\neach Chinese character at ~55px wide and overflowed the frame, while\n`Fontsize=12` rendered at ~30–35px wide and fit cleanly with 15-char\nlines.\n\nRule of thumb: start at `Fontsize=12`, render, then **always**\nextract a frame and look:\n\n```bash\n$FF -ss 30 -i output.mp4 -frames:v 1 /tmp/frame.png -y\n# then Read /tmp/frame.png to verify the longest-line cue fits\n```\n\nPick a timestamp that lands on the cue with the most characters per\nline — short lines won't expose overflow. Add `MarginL=20 MarginR=20`\nas a safety inset; never trust default left/right margins.\n\n#### Style cheatsheet\n\nKeys that matter (libass `force_style`):\n\n- `Fontname=PingFang SC` — macOS default CJK; alternates: `Songti SC`,\n  `Heiti SC`, `STHeiti`, `Hiragino Sans GB`.\n- `Fontsize=12` — start small, scale up only after frame check.\n- `PrimaryColour=&H00FFFFFF` — white text (BBGGRR + alpha).\n- `OutlineColour=&H00000000` — black outline.\n- `BorderStyle=1` — outline only (clean over varied backgrounds).\n  Use `BorderStyle=3` for an opaque box behind text when the\n  background is busy.\n- `Outline=2` — 2px outline thickness.\n- `Shadow=1` — subtle drop shadow.\n- `MarginL=20 MarginR=20` — keep text inside the frame.\n- `MarginV=40` — vertical distance from the bottom edge.\n\n#### SRT line-length discipline for burn-in\n\nEven with correct `Fontsize`, lines that are too long will wrap or\noverflow. Keep each on-screen line ≤ ~15 Chinese characters. Use\nexplicit `\\n` line breaks inside the SRT block — do not rely on\nauto-wrapping. Two short lines beat one long one every time.\n\n---\n\n## Chinese Voice Dubbing\n\nWhen the user wants the video to actually **speak Chinese** (not just\ndisplay Chinese subtitles), generate a TTS dub aligned to the original\nSRT timing.\n\n### Engine choice — Volcano (preferred) or edge-tts (fallback)\n\nFor Mandarin dubbing, the preferred engine is **Volcano (字节跳动豆包)\nTTS 2.0** — the voices are markedly more natural than edge-tts,\nespecially for emotional/contemplative content. It needs paid\ncredentials. Use edge-tts when you don't have Volcano access or while\ndebugging.\n\nThe bundled `dub.py` auto-routes by voice-ID prefix:\n\n- Voice starts with `zh_…_bigtts` → Volcano TTS 2.0\n- Voice starts with `zh-CN-…Neural` (or any locale + `…Neural`) →\n  edge-tts\n\n#### Volcano TTS — Chinese only\n\nEndpoint: `https://openspeech.bytedance.com/api/v3/tts/unidirectional`\n(used for both TTS 1.0 and 2.0; the Resource-Id header picks the\nbackend).\n\nHeaders:\n\n```\nX-Api-App-Id:       (env: VOLC_TTS_APPID)         # 10-digit speech App ID\nX-Api-Access-Key:   (env: VOLC_TTS_ACCESS_TOKEN)  # 32-char token from speech console\nX-Api-Resource-Id:  volc.service_type.10029       # see resource ID note below\nContent-Type:       application/json\n```\n\nLoading the credentials: most users keep them in `~/code/.env`. Read\nthem at the top of any session that needs them via:\n\n```bash\nset -a; source ~/code/.env; set +a\n```\n\n**Resource ID — important quirk.** The doc lists `seed-tts-2.0` as\nthe \"TTS 2.0 (recommended)\" resource, but a typical TTS-SeedTTS2.0\nconsole instance does **not** include the popular `*_bigtts` speaker\ncatalog (爽快斯斯, 高冷御姐, 开朗姐姐, etc.). Trying those speakers\nagainst `seed-tts-2.0` returns `200 code=55000000 \"resource ID is\nmismatched with speaker related resource\"`. The fix is to use\n`volc.service_type.10029` (the TTS 1.0 V3 endpoint) — the audio\nquality of the bigtts speakers is identical, and they all work\nagainst this resource. The bundled `dub.py` defaults to\n`volc.service_type.10029`; override with `VOLC_TTS_RESOURCE` env if\nyou have a different instance.\n\nOther 401/403 errors:\n\n- `401 code=45000010 \"load grant: requested grant not found in SaaS\n  storage\"` — the App ID + key combo is valid against the gateway,\n  but the user has not activated this resource. They must go to\n  火山引擎 → 语音技术 → 语音合成大模型 → 实例管理 and 开通 the\n  service. No workaround.\n- `403 code=45000030` — the speaker isn't included in the user's\n  instance bundle.\n\n**Response format.** Despite the doc's casual language, the response\nis **streaming NDJSON**, not a single JSON object and not raw audio\nbytes. Each line is a separate JSON event with a base64-encoded MP3\nchunk in `data`. The terminal event has `code: 20000000` (which\nmeans OK in this API's success codes — different from `code: 0`).\nConcatenate the decoded chunks for the full MP3.\n\n```python\nimport base64, json, requests\naudio = b\"\"\nr = requests.post(url, headers=h, json=payload, timeout=60, stream=True)\nfor line in r.iter_lines():\n    if not line: continue\n    evt = json.loads(line)\n    if evt.get(\"code\") not in (0, None, 20000000):\n        raise RuntimeError(f\"code={evt.get('code')} {evt.get('message')}\")\n    if evt.get(\"data\"):\n        audio += base64.b64decode(evt[\"data\"])\n```\n\n**Speaker catalog (verified working under\n`volc.service_type.10029`).** Full list at\nvolcengine.com/docs/6561/1257544 — but availability depends on your\ninstance bundle. Confirmed-working female voices for the typical\nSeedTTS-2.0 starter instance:\n\n| Speaker ID                                    | 中文名     | Feel                       |\n| ---                                           | ---        | ---                        |\n| `zh_female_gaolengyujie_moon_bigtts`          | 高冷御姐   | **Best for contemplative/spiritual content.** Mature, restrained, calm. |\n| `zh_female_kailangjiejie_moon_bigtts`         | 开朗姐姐   | Warm older-sister storytelling. |\n| `zh_female_shuangkuaisisi_moon_bigtts`        | 爽快斯斯   | Versatile, conversational baseline. |\n| `zh_female_linjianvhai_moon_bigtts`           | 邻家女孩   | Casual, lifestyle-vlog. |\n| `zh_female_yuanqinvyou_moon_bigtts`           | 元气女友   | Lively, upbeat. |\n| `zh_female_meilinvyou_moon_bigtts`            | 美丽女友   | Soft, intimate. |\n| `zh_female_shuangkuaisisi_emo_v2_mars_bigtts` | 斯斯情感版 | Full emotional range — pair with explicit emotion + scale. |\n\nThese voices return 55000000 against the typical instance even though\nthe doc lists them: `vv_uranus_bigtts`, `wenroushunv_moon_bigtts`,\n`qingxin_moon_bigtts`, `yingmaoxiaoyuan_moon_bigtts`,\n`tianxinxiaoling_moon_bigtts`, `shaoergushi_moon_bigtts`. Don't\npromise them without testing.\n\n**Audio params.** `speech_rate` is Volcano's native scale [-50, +100]\nwhere the value is a percentage delta (so `-8` means 8% slower). The\nscript passes `--rate -8%` through as `-8`. Useful emotion presets:\n\n- `emotion=\"calm\"`, `emotion_scale=4` — contemplative, default for\n  this skill's spiritual-content niche.\n- `emotion=\"gentle\"` — softer / more intimate.\n- `emotion=\"neutral\"` — flat / informational.\n- `emotion=\"sad\"` — melancholic. Use sparingly.\n\nOverride the script's defaults with `VOLC_TTS_EMOTION` and\n`VOLC_TTS_EMOTION_SCALE` env vars without editing code.\n\n**No English Volcano voices** are wired up in this skill — for\nEnglish use edge-tts (next section). Volcano does have English\nspeakers (`en_male_*_bigtts`, `en_female_*_bigtts`) but they aren't\ntypically included in TTS-SeedTTS-2.0 starter instances and they\nwere not validated for this workflow. Add them by extending the\nvoice routing in `dub.py` once verified.\n\n#### edge-tts (Microsoft Edge neural TTS)\n\nFree, no API key, high-quality but less expressive than Volcano.\nInstall into a project venv — **do not** call it via `uvx` once per\nsegment. Each `uvx` invocation spawns a fresh Python process and the\nbing endpoint will rate-limit or RST the connection after a handful\nof rapid hits, breaking mid-render.\n\n```bash\nuv venv .venv\nuv pip install --python .venv/bin/python edge-tts\n```\n\nThen drive it from a single long-lived Python process using\n`edge_tts.Communicate(...)` directly, with retry-on-failure logic. A\nready-to-run script lives at `scripts/dub.py` next to this SKILL.md;\ncopy it into the working directory and run:\n\n🛑 **Checkpoint — sample before full dub.** A full-video dub is the\nmost expensive step (TTS API calls + atempo + ffmpeg mux). Before\nrunning `dub.py` over the whole SRT:\n\n1. Pick the longest-text cue (worst stretch case) and one\n   short/casual cue (timbre check).\n2. Synthesize 3–4 voice/rate/pitch combos at 3–8s each — see \"Always\n   sample before committing\" below.\n3. Show the user the audio panel and ask: \"选哪个 voice?rate/pitch\n   要调吗?确认后我再跑全片。\" Wait for explicit pick.\n\nSkip the checkpoint only if the user named a specific voice up front\nAND has already heard a sample of that voice on this video.\n\n```bash\n.venv/bin/python dub.py [voice] [rate] [pitch]\n# e.g. mature warm female, slower, lower pitch:\n.venv/bin/python dub.py zh-CN-XiaoxiaoNeural -8% -10Hz\n```\n\nThe script:\n\n1. Reads the SRT (looks for `*.zh-CN.srt`-style filename — edit the\n   constants at the top of the script).\n2. Synthesizes one MP3 per cue under `dub_work/seg_NN.mp3`.\n3. Probes each clip's actual duration with `ffprobe`.\n4. For each cue: if TTS is longer than the SRT slot, chains `atempo`\n   filters to speed it up; if shorter, pads with silence after.\n5. Inserts silence segments for SRT gaps and any trailing tail so the\n   output audio length exactly matches the source video.\n6. Muxes the new audio into a `*_zh_dub.mp4` keeping the original\n   video stream by `-c:v copy`.\n\n### Voice selection — match the original speaker\n\nThis is the part that \"尽量匹配原声\" hinges on. There is no perfect\nmatch across language — choose gender, age feel, and tone\ndeliberately, then bend with rate/pitch.\n\n#### Chinese voices (Volcano preferred, edge-tts fallback)\n\nVolcano's `zh_female_gaolengyujie_moon_bigtts` (高冷御姐, calm,\n`speech_rate=-8`) is the validated baseline for mature contemplative\nfemale speakers — equivalent to or better than any edge-tts option\nfor that profile. See the Volcano speaker table above for the rest.\n\nIf Volcano isn't available, edge-tts catalog:\n\n| Voice                              | Gender | Default feel                  |\n| ---                                | ---    | ---                           |\n| `zh-CN-XiaoxiaoNeural`             | F      | Warm, news/novel              |\n| `zh-CN-XiaoyiNeural`               | F      | Lively, young                 |\n| `zh-CN-YunjianNeural`              | M      | Passionate, sports            |\n| `zh-CN-YunxiNeural`                | M      | Sunshine, lively              |\n| `zh-CN-YunyangNeural`              | M      | Professional newsreader       |\n| `zh-HK-HiuMaanNeural`              | F      | Friendly, slightly mature     |\n\n#### English voices (edge-tts neural, all multilingual)\n\nFor English dubbing, use edge-tts. All voices below speak fluent\nAmerican/British/Australian English; the `*Multilingual*` ones also\nhandle Spanish names, French/Italian loanwords, etc. without\nmispronunciation.\n\n| Voice                                  | Gender | Default feel                            |\n| ---                                    | ---    | ---                                     |\n| `en-US-AvaMultilingualNeural`          | F      | **Best for warm/mature/caring** — natural for spiritual or coaching content |\n| `en-US-EmmaMultilingualNeural`         | F      | Cheerful, conversational, younger        |\n| `en-US-AndrewMultilingualNeural`       | M      | Warm, confident, sincere                 |\n| `en-US-BrianMultilingualNeural`        | M      | Approachable, casual                     |\n| `en-US-AriaNeural`                     | F      | Crisp newsreader                         |\n| `en-US-GuyNeural`                      | M      | Steady male newsreader                   |\n| `en-GB-SoniaNeural`                    | F      | British female (RP)                      |\n| `en-GB-RyanNeural`                     | M      | British male (RP)                        |\n| `en-AU-WilliamMultilingualNeural`      | M      | Australian male                          |\n| `fr-FR-VivienneMultilingualNeural`     | F      | Mature European female who also reads English |\n\nFor matching a mature contemplative Spanish female (this skill's\ncanonical use case), start with `en-US-AvaMultilingualNeural` at\n`--rate -5% --pitch -3Hz`. Do **not** use the news-style `Aria` or\n`Guy` for spiritual content — they sound clinical.\n| `zh-TW-HsiaoChenNeural`            | F      | Friendly                      |\n| `en-US-AvaMultilingualNeural`      | F      | Caring, expressive (Western)  |\n| `fr-FR-VivienneMultilingualNeural` | F      | Mature European female        |\n\nPicking heuristics:\n\n- **Mature contemplative female speaker (yoga/spirituality/coaching):**\n  `zh-CN-XiaoxiaoNeural` with `--rate=-8% --pitch=-10Hz` is the most\n  reliable starting point — drops the youthful brightness, slows it\n  to a breathy pace.\n- **Mature professional male:** `zh-CN-YunyangNeural` with\n  `--rate=-5%`. Avoid Yunjian/Yunxi (too energetic).\n- **Young casual speaker:** Defaults; no pitch shift.\n- **Western-mouth feel** (sometimes preferred for foreign content):\n  one of the `*MultilingualNeural` voices. They're trained on\n  multiple languages and often have a less \"TV anchor\" timbre.\n\n### Always sample before committing\n\nDon't render the full dub on a guess. Generate the **same cue** with\n3–4 voice/parameter combos at 3–8 seconds each, and have the user pick\nA/B/C/D. Keeping samples under 10 seconds each lets the user audition\nthe entire panel in under a minute. After picking, re-run `dub.py`\nwith the chosen flags.\n\nThe skill's `scripts/sample_voices.py` (if present) is a thin wrapper\nthat does exactly this; otherwise just drive the same Python loop the\ndub script uses.\n\n### Filling awkward silences (Chinese is denser than Spanish)\n\nMandarin typically takes 60–80% of the time Spanish does to say the\nsame thing. With strict cue-by-cue timing, that leaves awkward 2–4s\nsilences at the end of most cues. Three levers, in increasing impact:\n\n1. **Slow the native TTS rate.** Changing `--rate` from `+0%` to\n   `-12%` to `-15%` produces clean, natural-sounding slower speech\n   (much better than time-stretching afterward). Try `-12%` first;\n   `-15%`/`-20%` for very contemplative content.\n\n2. **Mild slow-stretch per cue.** When a cue's TTS is still shorter\n   than its slot, run `atempo` between 0.82× and 0.95×. The bundled\n   `dub.py` does this automatically: when slack > 0.5s, it sets\n   `atempo = max(0.82, tts_dur / target_dur)` and pads the remainder.\n   Below 0.82× the voice starts sounding drugged; above 0.92× the\n   stretch is essentially imperceptible.\n\n3. **Expand the Chinese in the worst cues.** When the slot is so\n   long that even 0.82× stretch leaves >2s of silence, the cleanest\n   fix is to lengthen the translation. Add natural Mandarin\n   particles (\"嗯,\", \"其实\", \"也就是说\", \"你知道\") or unpack a\n   compressed phrase into its full meaning. This changes the\n   on-screen subtitle, so confirm with the user before doing it.\n   Edit the SRT, regenerate just those segments by deleting their\n   `dub_work/seg_NN.mp3` and re-running `dub.py`.\n\nCombine the levers: native rate `-12%` + stretch-to-fit handles ~80%\nof cases. Reserve text expansion for the 2–3 worst outliers.\n\n### Audio mixing — keep the original as a bed\n\nA pure dub-only track sounds dubbed (because it is). Mixing the\noriginal Spanish at low volume under the Chinese dub gives the\n\"professional translation\" feel — you still hear the speaker's breath,\nemphasis, and laughter, just under the Mandarin.\n\n```bash\n$FF -i original.mp4 -i dub.mp4 \\\n  -filter_complex \"[0:a]volume=0.18[orig];\\\n                   [1:a]volume=1.0[dub];\\\n                   [orig][dub]amix=inputs=2:duration=longest:normalize=0[a]\" \\\n  -map 0:v -map \"[a]\" \\\n  -c:v copy -c:a aac -b:a 192k mixed.mp4\n```\n\nReasonable starting volumes:\n\n- Spanish bed at `0.15`–`0.25` (≈ −16 to −12 dB)\n- Chinese dub at `1.0`\n- Use `normalize=0` so amix doesn't auto-attenuate when both are\n  active.\n\n### Combining dub + burn-in + bed (the full job)\n\nOne ffmpeg call does all three at once — burn the Chinese subtitle\nonto the video stream and mix the two audio tracks:\n\n```bash\n$FF -i original.mp4 -i dub.mp4 \\\n  -filter_complex \"[0:v]subtitles=input.zh-CN.srt:force_style='Fontname=PingFang SC\\,Fontsize=12\\,PrimaryColour=&H00FFFFFF\\,OutlineColour=&H00000000\\,BorderStyle=1\\,Outline=2\\,Shadow=1\\,MarginL=20\\,MarginR=20\\,MarginV=40'[v];\\\n                   [0:a]volume=0.18[orig];[1:a]volume=1.0[dub];\\\n                   [orig][dub]amix=inputs=2:duration=longest:normalize=0[a]\" \\\n  -map \"[v]\" -map \"[a]\" \\\n  -c:v libx264 -crf 18 -preset medium -pix_fmt yuv420p \\\n  -c:a aac -b:a 192k final.mp4\n```\n\nThis is the \"ship to social media\" final cut.\n\n---\n\n## File Naming\n\nDerive everything from the source video's stem. BCP-47-style suffixes\nmake the target language obvious at a glance and keep multiple\ntarget-language outputs side-by-side.\n\n```text\ninput:                       entrevista.mp4\n\nChinese pipeline:\n  whisper Spanish SRT        entrevista.srt\n  Chinese SRT                entrevista.zh-CN.srt\n  Chinese dub (audio only)   entrevista_zh_dub.mp4\n  Chinese final (subs+dub)   entrevista_zh_final.mp4\n\nEnglish pipeline:\n  English SRT                entrevista.en.srt\n  English dub                entrevista_en_dub.mp4\n  English final              entrevista_en_final.mp4\n\nBilingual subtitles:\n  Spanish + Chinese          entrevista.es-zh.srt\n  Spanish + English          entrevista.es-en.srt\n  three-language             entrevista.es-zh-en.srt\n```\n\nWhen only one target language is in play, `*_final.mp4` without a\nlanguage tag is acceptable (matches the existing project examples).\nAdopt the `_zh_` / `_en_` infix the moment a project produces both.\n\n---\n\n## Important Constraints\n\n- Do not hallucinate unclear Spanish.\n- Do not summarize unless the user asks.\n- Do not remove important details.\n- Do not invent timestamps.\n- Do not change the meaning to make it sound smoother.\n- Do not over-explain inside subtitles.\n- Do not translate proper nouns incorrectly.\n- Do not use Traditional Chinese unless requested.\n- Do not output invalid SRT syntax.\n- Do not include markdown code fences inside the final `.srt` file if creating an actual file.\n- Do not promise burn-in without first verifying the available\n  `ffmpeg` was built with libass. If it wasn't, switch to soft-mux or\n  pull a static build to `/tmp` rather than reinstalling system\n  ffmpeg.\n- Do not commit a burned-in render without extracting at least one\n  frame at a long-line cue and visually confirming text is fully on\n  screen.\n- Do not dub by calling `uvx edge-tts` per cue — use the persistent\n  Python library path with retries or the API will RST mid-render.\n- Mild slow-stretch (`atempo` between **0.82× and 0.95×**) is fine and\n  often necessary — both Mandarin and English tend to finish faster\n  than Spanish, leaving awkward dead air. English may be marginally\n  shorter than Spanish (~85%); Chinese is much shorter (~65%). Cap\n  at 0.82×; below that the voice sounds drugged. Do not stretch when\n  the slack is under ~0.5s (a short tail pause sounds natural).\n- Don't promise a Volcano voice without checking it works against the\n  user's instance. The doc lists many voices that error with\n  `code=55000000 \"resource ID is mismatched with speaker related\n  resource\"` against typical SeedTTS-2.0 starter bundles. The skill's\n  Volcano speaker table marks which are confirmed-working. **Mandatory\n  smoke test before promising any Volcano voice on a new account:**\n  synth one ~5-word cue with that speaker ID first; only quote it to\n  the user if the smoke test returns a non-empty MP3. If the smoke\n  test 401s with `code=45000010` (\"grant not found\"), tell the user\n  they need to 开通 the resource in 火山引擎 console — do not pretend\n  it'll work after a retry.\n- For the Volcano endpoint, parse the response as **streaming\n  NDJSON**, not a single JSON document. The success terminator is\n  `code=20000000`, not `code=0`. Concatenate base64-decoded `data`\n  chunks for the full MP3.\n- **Default to one voice for the whole video.** Do not run\n  diarization, do not tag cues with `[A]`/`[B]`, and do not bring\n  up multi-speaker complexity unless the user explicitly says the\n  source has multiple speakers. Adding speaker separation later is\n  a quick re-run; shipping confused multi-voice output by default\n  wastes the user's time. The \"Advanced: Multi-speaker dubbing\"\n  section is opt-in only.\n\n---\n\n## Advanced: Multi-speaker dubbing (opt-in)\n\n**Only invoke this section when the user explicitly says the source\nhas multiple speakers** (\"interview\", \"two people\", \"dialogue\",\n\"separate the speakers\", \"different voice for each\", or a direct\nrequest to do diarization). For everything else, use one voice and\nskip this section.\n\nWhen triggered, generate the dub with a different voice per speaker\nso the listener can follow who's speaking. Two paths:\n\n### Path 1 (recommended for on-camera speakers): visual diarization\n\n`scripts/visual_diarize.py` watches mouth movement per face per\nframe and tags each cue with the dominant speaker. Self-contained,\nno API keys, no audio fingerprinting.\n\n```bash\nuv pip install --python .venv/bin/python mediapipe opencv-python\n\n.venv/bin/python scripts/visual_diarize.py \\\n    --video input.mp4 --srt input.en.srt \\\n    --out input.en.diarized.srt \\\n    --report diarization_report.json \\\n    --sample-fps 5 --num-speakers 2\n```\n\nHow it works:\n\n1. Samples N frames per second (default 5).\n2. Runs MediaPipe FaceLandmarker (Tasks API) for up to\n   `--num-speakers` faces per frame, 478 landmarks each.\n3. Measures mouth aperture per face as the vertical distance between\n   inner upper lip (idx 13) and inner lower lip (idx 14).\n4. Bins faces by horizontal screen position (x-quantiles) → speakers\n   `A`, `B`, ... left-to-right.\n5. For every cue's [start, end] window, integrates per-speaker\n   frame-to-frame mouth-aperture change. Highest mover wins the tag.\n6. Writes a `[A]`/`[B]`-prefixed SRT plus a JSON report with\n   per-cue scores and a confidence ratio (winner / runner-up).\n\nOn first run, downloads the FaceLandmarker model (~3.6 MB) to\n`/tmp/mp_models/face_landmarker.task`.\n\n**Visual is materially better than guessing from text.** In one\nvalidation, manual text-based labels split 6/50 between speakers;\nvisual diarization showed the actual split was 29/27 — text-based\nguessing was wildly wrong because both people take similar-shaped\nturns. Always prefer visual when the speakers are on camera.\n\n**Spot-check low-confidence cues.** Any cue in the JSON report with\n`confidence_ratio < 1.5` is borderline — usually overlapping speech\nor one speaker briefly off-frame. Hand-correct before dubbing.\n\n### Path 2 (fallback): manual tagging\n\nFor very short clips (1–2 minutes), or when speakers are off-camera,\nor when visual diarization fails:\n\n```text\n1\n00:00:00,000 --> 00:00:03,400\n[A] So what about that AI rewrite thing?\n\n2\n00:00:03,400 --> 00:00:08,200\n[B] Right — let me explain the workflow.\n```\n\nSave as `*.tagged.srt`. Keep the **clean** SRT (without tags) for\nsubtitle burn-in.\n\n### Routing voices in dub.py\n\nPass `--voice-map` with `speaker=voice` pairs. The positional voice\narg is the default for cues with no tag.\n\n```bash\n.venv/bin/python scripts/dub.py en-US-AndrewMultilingualNeural -3% +0Hz \\\n    --srt input.en.tagged.srt \\\n    --voice-map \"A=en-US-BrianMultilingualNeural,B=en-US-AndrewMultilingualNeural\"\n```\n\nVoice-pairing tips:\n\n- **Two of the same gender:** pick voices with audibly different\n  timbre. Brian (casual) + Andrew (warm) works for two American\n  males. Ava (warm female) + Emma (cheerful female) for two females.\n- **Mixed gender:** Ava + Andrew is a clean default.\n- **Accent contrast:** pair `en-US-` and `en-GB-` for distinctness.\n- **Chinese:** mix Volcano voices like\n  `zh_female_gaolengyujie_moon_bigtts` (mature) +\n  `zh_female_kailangjiejie_moon_bigtts` (warm sister).\n\n### Rendering with multi-speaker dub\n\nUse the **clean** SRT for `render.py --srt` so `[A]`/`[B]` tags\ndon't appear on screen:\n\n```bash\n.venv/bin/python scripts/render.py \\\n    --video input.mp4 \\\n    --srt input.en.srt              \\  # CLEAN, for burn-in\n    --dub input_en_dub.mp4          \\  # contains the per-speaker voices\n    --out input_en_final.mp4\n```\n\n### Limits\n\nVisual diarization fails when:\n\n- A speaker is consistently off-camera while talking.\n- Camera cuts or zooms make face position unstable across cues.\n- Three or more speakers sit at similar horizontal positions\n  (x-quantile binning is too coarse — switch to k-means on (x, y)\n  or use audio-based diarization instead).\n\nFor audio-only material (podcasts, voice-overs), fall back to\n`pyannote.audio` or `whisperx --diarize`. This skill does not yet\nbundle audio-based diarization — file an issue or PR if you need it.\n\n---\n\n## Default Final Response\n\nWhen the task is complete, respond briefly. Match the user's language\nand use the target-language tag in the summary.\n\nFor Chinese:\n\n```text\n已完成:\n- 西班牙语转写\n- 中文翻译\n- 中文字幕 SRT 文件\n(如有:中文配音、烧入字幕、原声底层)\n\n不确定片段:\n- 00:01:23–00:01:26 背景噪音较大,原文可能不完全准确。\n```\n\nFor English:\n\n```text\nDone:\n- Spanish transcript\n- English translation\n- English subtitle SRT\n  (plus: English dub, burned-in subs, Spanish bed if produced)\n\nUncertain segments:\n- 00:01:23–00:01:26 background noise; source partially unclear.\n```\n\nIf there are no uncertain parts, drop the second list.\n",
  "last_post_text": "translate-video skill 更新 — 加了三道用户确认门, 防止误触烧 CPU/烧 API:\n\n🛑 烧入字幕前: 先渲 30s 预览 + 抽帧 + 你确认字号才跑全片\n🛑 全片配音前: 强制 sample 3-4 个 voice 让你挑\n🛑 承诺 Volcano 声音前: 5 词 smoke test, 401 直接提示去开通\n\nhttps://github.com/jianshuo/translate-video\n#ClaudeCode",
  "last_tweet_id": "2053491380497031580",
  "last_tweet_url": "https://x.com/jianshuo/status/2053491380497031580"
}
writing-skillsobra/superpowersUse when creating new skills, editing existing skills, or verifying skills work before deploymentStars: 323kUpdated Claude CodeCodexVerifiedInstall
xlsxanthropics/skillsUse this skill any time a spreadsheet file is the primary input or output. This means any task where the user wants to: open, read, edit, or fix an existing .xlsx, .xlsm, .xltx, .csv, or .tsv file (e.g., adding columns, computing formulas, formatting, charting, cleaning messy data); create a new spreadsheet from scratch or from other data sources; or convert between tabular file formats. Trigger especially when the user references a spreadsheet file by name or path — even casually (like "the xlsx in my downloads") — and wants something done to it or produced from it. Also trigger for cleaning or restructuring messy tabular data files (malformed rows, misplaced headers, junk data) into proper spreadsheets. The deliverable must be a spreadsheet file. Do NOT trigger when the primary deliverable is a Word document, HTML report, standalone Python script, database pipeline, or Google Sheets API integration, even if tabular data is involved.Stars: 198kUpdated Claude CodeCodexVerifiedInstall
shipping-and-launchaddyosmani/agent-skillsPrepares production launches. Use when preparing to deploy to production, or when asking what needs to be in place before shipping. Use when you need a pre-launch checklist, when setting up monitoring, when planning a staged rollout, or when you need a rollback strategy.Stars: 119kUpdated Claude CodeCodexVerifiedInstall
childhoodthedotmack/claude-memBabysit a shipped release after merge, publication, or deployment. Verify published packages, production behavior, installed plugins, and late CI or user reports until the observation window is clean; use babysit for PR readiness before release.Stars: 118kUpdated Claude CodeCodexVerifiedInstall