YouTube Slide Maker
YouTube URL → スライドデッキ or 切り抜き動画 の2モード。字幕 → LLMでキーフレーム選定 → ffmpegで抽出。
Modes
| Mode | Output | Use Case |
|---|
| slides | slides.md (Marp) + frames/*.png | 動画の要点スライド化（PKM・ブログ等） |
| clips | clips/*.mp4 (1-3分の切り抜き) | SNS・ダイジェスト用 |
| both | 両方 | スライドと切り抜きを同時生成 |
Workflow
1. Prepare
# Default output directory (PKMのInbox配下)
OUT="Projects/PKM/youtube-slides/<video-slug>"
mkdir -p "$OUT/frames" "$OUT/clips"
cd "$OUT"
2. Download video + subtitles
# Video (prefer 720p to save space)
yt-dlp -f 'bv[height<=720]+ba/b[height<=720]' \
--merge-output-format mp4 \
--write-auto-subs --write-subs --sub-lang en,ja --sub-format vtt \
--convert-subs srt \
-o 'video.%(ext)s' "$URL"
--merge-output-format mp4 を付けること。付けないと結合結果が video.webm /
video.mkv になり、以降の video.mp4 決め打ちの ffmpeg コマンドが全部落ちる。
それでも拡張子が変わることがあるので、続くコマンドは実ファイルを拾って使う:
VIDEO=$(ls video.mp4 video.mkv video.webm 2>/dev/null | head -1)
Fields of interest after download:
$VIDEO（通常 video.mp4） — full videovideo.en.srt or video.ja.srt — transcript
3. Analyze transcript → select keyframes
Read the SRT. For a target of N slides (default 10-15):
- Identify chapter-like structure (topic shifts, emphasis words, demos)
- Output JSON-like list:
[{ts: "00:01:23", title: "...", summary: "..."}]
- For clips mode: identify 1-3 min contiguous segments (
{start, end, title})
Prompt pattern for transcript → keyframes:
This is a video transcript (SRT). Select {N} key moments that best represent the video.
For each, output:
- ts: timestamp in HH:MM:SS (pick a moment ~2-5s after speaker introduces topic to catch the explanation frame)
- title: ≤30 char headline
- summary: 1-2 sentence body
4. Extract frames
# Per keyframe (ts = HH:MM:SS)
ffmpeg -ss "$ts" -i "$VIDEO" -vframes 1 -q:v 2 "frames/$(printf '%02d' $i)-$slug.png"
Tips:
-ss before -i = fast seek (input-level)-q:v 2 = high quality JPEG (or omit for PNG)- For slide thumbnails: scale with
-vf scale=1280:-1
5. Extract clips (clips/both mode)
# Lossless cut (fast, no re-encode)
ffmpeg -ss "$start" -to "$end" -i "$VIDEO" -c copy "clips/$(printf '%02d' $i)-$slug.mp4"
# Re-encode (needed if keyframe alignment matters, smooth cut)
ffmpeg -ss "$start" -to "$end" -i "$VIDEO" -c:v libx264 -c:a aac "clips/$(printf '%02d' $i)-$slug.mp4"
For SNS切り抜き: consider vertical crop
ffmpeg -i clip.mp4 -vf "crop=ih*9/16:ih,scale=1080:1920" -c:a copy vertical.mp4
6. Generate Marp slides (slides/both mode)
See references/marp-template.md for the slide structure.
Key points:
- 1 frame per slide, title overlay + summary bullet
- Use
Projects/PKM/youtube-slides/<slug>/slides.md
- Render to PDF/HTML:
npx @marp-team/marp-cli slides.md -o slides.pdf
7. Link back to PKM
Add to the original Readwise inbox file (Inbox/readwise-*.md) or Journal:
### Slide Deck
- slides
- Source: <video URL>
- Generated: YYYY-MM-DD
Parameters (ask if unclear)
- URL: required
- Mode:
slides | clips | both (default: slides)
- N: slide count (default: 12) or clip count (default: 3)
- Language: subtitle preference (
en, ja, auto)
- Output dir: default
Projects/PKM/youtube-slides/<slug>/
Common Pitfalls
- No subtitles available: fall back to
--write-auto-subs (auto-generated). If still none, use whisper or ask user.
- Slug conflicts: derive from video title + short ID hash
- Large videos: limit to 720p by default, warn if >1GB
- Rate limits: yt-dlp may need
--cookies-from-browser for members-only or age-restricted
- Timestamp offset:
-ss before -i is fast but can be ~1s off on some codecs; use -ss after -i for frame-accurate at cost of speed
Tool Reference
See references/commands.md for full yt-dlp / ffmpeg command cookbook.
Integration with PKM
- Output dir:
Projects/PKM/youtube-slides/<slug>/ (submodule-safe location)
- Slide deck reference added to source
Inbox/readwise-*.md file
- No Cards auto-creation — user decides if content warrants a Card