メニュー
Extract and process subtitles from a SOLAR LINE episode video. Supports YouTube VTT download, Whisper STT, video OCR, and dialogue line extraction.
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
Full pipeline for analyzing a new SOLAR LINE episode — from subtitle extraction through orbital analysis to published report. Use when starting analysis of a new episode.
Review a SOLAR LINE episode or summary report for quality, accuracy, and readability. Checks data integrity, Japanese text quality, and orbital mechanics correctness.
Consult OpenAI Codex as a trusted friend for design review, architecture feedback, or technical advice. Use when the user asks to consult Codex or wants a second opinion.
Analyze Claude Code session costs using ccusage. Shows token distribution, per-model breakdown, cache efficiency, and optimization recommendations.
SOC 職業分類に基づく
| name | subtitle-extraction |
| description | Extract and process subtitles from a SOLAR LINE episode video. Supports YouTube VTT download, Whisper STT, video OCR, and dialogue line extraction. |
| argument-hint | <video-id-or-path> --episode <N> |
Extract subtitles from SOLAR LINE episodes using multiple sources.
cd ts
npm run collect-subtitles -- <youtube-video-id> --lang ja --out-dir ../raw_data/subtitles
npm run extract-dialogue -- ../raw_data/subtitles/<vtt-file> --episode <N> --video-id <id>
cd ts
# Step 1: Download audio (run in background — takes ~1 min)
# yt-dlp -x --audio-format wav -o ../raw_data/audio/ep<N>.wav "https://youtu.be/<video-id>"
# Step 2: Run Whisper (run in background — takes 10-26 min CPU-only)
npm run whisper -- ../raw_data/audio/ep<N>.wav --model large-v3-turbo --language ja
# Step 3: Process output
npm run process-whisper -- ../raw_data/whisper/<output>.json --episode <N> --video-id <id>
# Step 4: Extract dialogue lines
npm run extract-dialogue-whisper -- ../raw_data/whisper/<processed>.json --episode <N>
cd ts
# Step 1: Extract keyframes from video
python src/extract-frames.py ../raw_data/<video-file> ../raw_data/frames/ep<N>
# Step 2: Run OCR on extracted frames
python src/video-ocr.py ../raw_data/frames/ep<N> --episode <N>
Output: reports/data/episodes/epNN_ocr.json
npm run compare-transcriptions -- ../raw_data/subtitles/<vtt> ../raw_data/whisper/<whisper> --episode <N>
| Step | Output | Location |
|---|---|---|
| YouTube VTT | Raw subtitle file | raw_data/subtitles/ (gitignored) |
| Whisper | Raw transcription JSON | raw_data/whisper/ (gitignored) |
| Video OCR | OCR data JSON | reports/data/episodes/epNN_ocr.json |
| Phase 1 Extract | epNN_lines.json | reports/data/episodes/ |
| Phase 2 Attribute | epNN_dialogue.json + epNN_speakers.json | reports/data/episodes/ |
run_in_background for yt-dlp and Whisper to avoid inflating session context| Source | Type Code | Accuracy (EP01) | Notes |
|---|---|---|---|
| YouTube auto-generated | youtube-auto | 68.3% | Default VTT download |
| YouTube manual | youtube-manual | — | Rare, higher quality |
| Whisper (large-v3-turbo) | whisper | 91.4% | Local inference, best accuracy |
| Whisper (medium) | whisper | 82.6% | Faster, lower accuracy |
| Video OCR (Tesseract) | video-ocr | 10.1% | On-screen text only |
| Manual | manual | — | Hand-transcribed |