SuperAPI AI Directory
Partner with us中文AI newsFavorites
Skillswjs-creating-video-book

wjs-creating-video-book

Design & creation

View repository

Updated 10-15 23:32

About this skill

wjs-creating-video-book

Turn a book into chapter-by-chapter YouTube videos: each chapter uses a VoiceDrop audiobook mp3 as the audio track, distill 1-3 central ideasfrom the chapter content, design a prompt for each central idea, generate visuals with GPT Image 2 (Ken Burns slow push + crossfade), overlay the central idea as large text on the visuals, and finally composite a 1920×1080 landscape video and upload it to YouTube.

What this skill produces

Dimension Default
Size 1920×1080 landscape (16:9), 30fps
Granularity One video per chapter(the video is as long as the chapter mp3)
Audio track VoiceDrop chapter audiobook mp3 (no re-TTS)
Visuals 1-3 GPT Image 2 generated images per chapter, Ken Burns slow push, 1s crossfade between segments
Text For each central idea: large title (≤12 characters) + one explanatory sentence, fade in/out, rendered with Pillow
Project directory ~/code/book-videos/<book-slug>/
Output <slug>-chNN.mp4 × N chapters + UPLOAD_META.md
Publish wjs-uploading-video upload (landscape regular video), default public

When this skill fires

  • The user gives a book and says "make it into a video" or "add visuals to the audiobook and post it to YouTube"
  • The user runs /wjs-creating-video-book <书名或书架章节>

When NOT to use

  • What they want is a 30-90 second vertical explainer short → /wjs-converting-text-to-video
  • They only want the audiobook mp3, no video → /wjs-voicedrop-reading-aloud
  • They want to write a book review article → /wjs-publishing-wechat

Core Principle

This is a "listening-first, watching-second" video. The audio track is the full-chapter narration, and the audience is "listening to a book"; the job of the visuals is to give the eyes an anchor to rest on, not to put on an animated performance. So:

  • Fewer central ideas rather than more — pick 1-3 true pillar points per chapter; it is normal for one image to stay on screen for several minutes
  • One visual style for the whole book — the art style must be consistent from chapter to chapter (same medium, same color family), so the audience isn't pulled out of the experience when switching chapters
  • Draw concrete scenes, not abstract concepts — for "leverage", draw a long pole prying up a boulder, not the word "leverage" or an abstract arrow chart
  • Copyright self-check — narrating the full original text of a chapter + public release only applies to public-domain books or books the user has rights to; for copyrighted books, confirm with the user before publishing

Workflow

Step 0: Project directory

SLUG=<book-slug>   # 优先用书的英文名短 slug（如 naval-almanack）；没有英文名才用拼音
BOOK=~/code/book-videos/$SLUG
mkdir -p $BOOK/chapters/{01,02,...}   # 每章一个两位数目录

Step 1: Get the text + audiobook mp3 for each chapter

Material Source (by priority)
Chapter mp3 ① Download the already-composited chapter mp3 from the jianshuo.dev bookshelf audiobook (R2, see memory [[jianshuo-dev-audiobook]]) ② If not available → use /wjs-voicedrop-reading-aloud to generate from the chapter text
Chapter text Bookshelf page / PDF or EPUB provided by the user (for EPUB first pandoc x.epub -t plain, if pandoc isn't installed first brew install pandoc）

Save as chapters/NN/audio.mp3 and chapters/NN/chapter.md。

Step 2: Distill 1-3 central ideas per chapter → ideas.json

Read chapter.md, pick the truly load-bearing ideas. The number follows audio length: 1-2 for <5 minutes, 2-3 for 5-15 minutes, and for >15 minutes alsoat most 3(the visuals are anchors, not subtitles).

[
  {"title": "杠杆是新的钱",  "caption": "代码和媒体是无需许可的杠杆", "image": "img-1.png", "start": 0},
  {"title": "把自己产品化",  "caption": "独特知识乘以杠杆才有复利",   "image": "img-2.png", "start": 312}
]
  • title ≤12 characters (large text);caption one sentence ≤20 characters; neither ends with punctuation
  • start(seconds) = the position where that idea starts being spoken in the narration; estimating by text position ratio is fine; if all are omitted, divide the audio evenly
  • The central idea must stay faithful to the chapter content — this is not a derivative work, it is highlighting key points for the listener

Step 3: One visual per central idea (GPT Image 2)

First givethe whole booka style base draft $BOOK/style.md(do it only once):the medium must be illustration(editorial illustration / watercolor illustration / minimalist flat illustration, etc.; do not use photorealism — illustration holds up best under Ken Burns slow push-in and text overlay), color family, lighting mood, 2-3 sentences.

Assemble each image's prompt according to thevideo-friendly illustration template= whole-book illustration style + a concrete scene for that idea + fixed ending:

node ~/.claude/skills/gpt-image-2-skill/scripts/gpt_image_2_skill.cjs \
  --json --provider codex images generate \
  --prompt "<全书插画风格>. <这个中心思想的具象场景>. Wide horizontal panoramic composition, landscape orientation, cinematic 16:9 framing. Generous negative space, calm lower left area for text overlay. Illustration, not a photograph. No text, no words, no letters." \
  --out $BOOK/chapters/NN/img-1.png --format png --size 1920x1088 --quality high

The fixed ending's three sentences are all indispensable; each fixes a pitfall actually encountered:

  • "Wide horizontal panoramic composition, landscape orientation" — the Codex side --size is only a reference; the composition actually follows the prompt; if you do not write this sentence, it returns a square image or even a vertical image (after a vertical image is center-cropped to 16:9, most of the frame is lost). Although the rendering script has fallback cropping, getting a horizontal image at the source is the right solution
  • "Generous negative space, calm lower left area" — the lower-left corner is the fixed position for the large central-idea text, and the main subject of the image must avoid it
  • "No text, no words, no letters" + "Illustration, not a photograph" — the text is overlaid by the rendering script with Pillow; AI drawing Chinese text will definitely fail; photorealism looks cheap under slow push-in + large text

Other requirements:

  • Draw concrete scenes (people, objects, light, space), not conceptual diagrams / charts / arrows
  • Multiple images in the same chapter must have visual connections (different angles of the same space, different times of the same day)
  • Check the dimensions after generation: if it still produces a vertical image (height > width), regenerate with more explicit horizontal-scene wording; do not settle

Step 4: Synthesize the chapter video

cd $BOOK/chapters/NN
uvx --with pillow python ~/.claude/skills/wjs-creating-video-book/scripts/render-chapter.py .
mv chapter.mp4 $BOOK/$SLUG-chNN.mp4

What the script (already verified on this machine) does: after upsampling each image to 3840, zoompan slow push-in (about 1.09× for 12s), 1s xfade between segments, render the large central-idea text + explanatory sentence with Pillow into a transparent PNG and then overlay it (1s alpha fade in/out, avoiding 0.8s/0.5s at the start and end of each segment), use the chapter mp3 directly as the audio track,-crf 19 output. The font is auto-detected as PingFang → Hiragino Sans GB → STHeiti.

Note: ffmpeg on this machine was not compiled with drawtext, so text must go through the Pillow path in the script; do not try to hand-write a drawtext filter.

Step 5: Thumbnail (per chapter)

Use a video frame + large text (do not draw another AI illustration): pick the frame where the chapter title text is clearest:

ffmpeg -y -ss <标题完全显示的秒数> -i $BOOK/$SLUG-chNN.mp4 -frames:v 1 -vf scale=1280:720 $BOOK/thumb-chNN.jpg

uploader does not support setting thumbnails via API — after uploading, remind the user to set it manually in YouTube Studio.

Step 6: UPLOAD_META.md + upload

The format must strictly copy the skeleton below — upload_youtube.py the parser only recognizes ## NN · 文件名 block headers + **短标题** / **视频描述** sections + #tagin the body; other formats parse to 0 blocks and exit directly. One block per chapter:

## 01 · <slug>-ch01.mp4

**短标题**
《书名》第一章：<章名或一句话钩子>

**视频描述**
<本章讲什么，2-3 句> + 本章的中心思想列表

有声书朗读来自 VoiceDrop。

#有声书 #听书 #<书名> #<作者> #读书

---

uvx --with google-auth --with requests python \
  ~/.claude/skills/wjs-uploading-video/scripts/upload_youtube.py --dir $BOOK

When there are many chapters, pay attention to the YouTube daily quota (about 6/day by default): create a upload-batch-N/ staging directory to hold soft links to the mp4s to upload that day + UPLOAD_META.md soft links,--dir point to the staging directory and upload day by day.

--dir mode automatically pairs the mp4 in the directory with the UPLOAD_META.md（Do not use --video + --meta — --meta only takes effect on the --dir branch,--video branch requires --title and ignores meta files). For multi-chapter videos, it is recommended to add --playlist <id> to group them into the same playlist. After uploading, send the link back to the user.

Step 7: Attach the full transcript as comments

After each video is uploaded, attach that chapter's transcript below the video in the form of a "pinned comment + replies split by section" (people listening to the audiobook can read along):

uvx --with google-auth --with requests python \
  ~/.claude/skills/wjs-creating-video-book/scripts/post-transcript-comments.py \
  <video_id> chapters/NN/chapter.md \
  --header "📖《书名》第 N 章「章名」全文文稿"

Script behavior: automatically strips the site header and footer from chapter.md, splits by ## sections into 3-6 comments (the first is the pinned main comment, the rest are replies to it, each ≤4200 characters, labeled with k/N numbering), with a 2s interval between replies to prevent spam detection.

Permission: requires the youtube.force-ssl scope. If the token only has youtube.upload(posting comments returns 403 / the script prompts re-auth), run uvx --with google-auth-oauthlib python ~/.claude/skills/wjs-creating-video-book/scripts/reauth-youtube.py once — it will pop up a browser for the user to click authorize once, and the new token will contain both upload + force-ssl, so it won't pop up again in the future.

Directory structure

~/code/book-videos/<slug>/
├── style.md                  # 全书视觉风格底稿（一次定）
├── <slug>-ch01.mp4 ...       # ⭐ 每章最终视频（在根目录，供 --dir 上传扫描）
├── thumb-ch01.jpg ...        # 每章缩略图（YouTube Studio 手动设）
├── UPLOAD_META.md
└── chapters/
    └── 01/
        ├── chapter.md        # 章节文本
        ├── audio.mp3         # VoiceDrop 读书 mp3
        ├── ideas.json        # 1-3 个中心思想
        ├── img-1.png ...     # GPT Image 2 画面
        └── overlay-1.png ... # 脚本生成的文字层（中间产物）

Anti-Patterns

Do not Reason
Re-TTS the chapter text The audio track is the VoiceDrop reading mp3; re-synthesizing is both wasteful and inconsistent
Cram 5-6 central ideas into one chapter The visuals become a subtitle machine. Fewer anchors are better than more, ≤3
Have the AI draw text in the prompt GPT Image 2 always breaks when drawing Chinese. Force "No text" in the prompt, and overlay text via Pillow
Not writing landscape/wide composition in the prompt On the Codex side, --size is only a reference; it will return square or portrait images, and 16:9 cropping loses most of the frame
Use a photo-realistic style Only illustration can withstand Ken Burns slow push + large text overlay; the photo style looks cheap
Draw abstract concept diagrams (arrows/charts/concept words) Draw concrete scenes and let the visuals breathe on their own
Change the art style for each chapter One style.md for the whole book, consistent across chapters
Hand-write the drawtext filter The local ffmpeg was not compiled with drawtext, so it will say "No such filter". Use the script's Pillow path
Publish copyrighted books directly and publicly The copyright risk of reading the full original text of an entire chapter is far greater than that of a book review. Only publish by default for public-domain books / books you have rights to; otherwise confirm first
Draw a separate AI illustration for the thumbnail Use a video frame + large text (same cover convention)
For upload use --video + --meta This combination does not hold (--video requires --title and ignores meta), use --dir

Dependencies

  • ffmpeg(already installed; note that this machine's build has no drawtext, and the script already bypasses it with Pillow)
  • uvx + pillow — text layer rendering (uvx --with pillow python ...）
  • GPT Image 2（~/.claude/skills/gpt-image-2-skill/，--provider codex) — image generation
  • wjs-voicedrop-reading-aloud — generated when the chapter mp3 does not exist
  • wjs-uploading-video + OAuth token ~/.config/youtube/token.json — upload
  • pandoc — only needed for EPUB input, not installed by default on this machine, first brew install pandoc