YouTube Transcript

Downloads transcripts (subtitles/captions) from YouTube videos. Works with both manually created and auto-generated transcripts. No API key or browser required — uses YouTube's InnerTube API directly and automatically falls back to yt-dlp when YouTube blocks the direct API path.

Fetches video metadata and cover image on first run, caches raw data for fast re-formatting.

Script Directory

Scripts in scripts/ subdirectory. {baseDir} = this SKILL.md's directory path. Resolve ${BUN_X} runtime: if bun installed → bun; if npx available → npx -y bun; else suggest installing bun. Replace {baseDir} and ${BUN_X} with actual values.

Usage

# Default: markdown with timestamps (English)
${BUN_X} {baseDir}/scripts/main.ts \x3Cyoutube-url-or-id>

# Specify languages (priority order)
${BUN_X} {baseDir}/scripts/main.ts \x3Curl> --languages zh,en,ja

# Without timestamps
${BUN_X} {baseDir}/scripts/main.ts \x3Curl> --no-timestamps

# With chapter segmentation
${BUN_X} {baseDir}/scripts/main.ts \x3Curl> --chapters

# With speaker identification (requires AI post-processing)
${BUN_X} {baseDir}/scripts/main.ts \x3Curl> --speakers

# SRT subtitle file
${BUN_X} {baseDir}/scripts/main.ts \x3Curl> --format srt

# Translate transcript
${BUN_X} {baseDir}/scripts/main.ts \x3Curl> --translate zh-Hans

# List available transcripts
${BUN_X} {baseDir}/scripts/main.ts \x3Curl> --list

# Force re-fetch (ignore cache)
${BUN_X} {baseDir}/scripts/main.ts \x3Curl> --refresh

Options

Optional Environment Variables

Input Formats

Accepts any of these as video input:

• Full URL: https://www.youtube.com/watch?v=dQw4w9WgXcQ
• Short URL: https://youtu.be/dQw4w9WgXcQ
• Embed URL: https://www.youtube.com/embed/dQw4w9WgXcQ
• Shorts URL: https://www.youtube.com/shorts/dQw4w9WgXcQ
• Video ID: dQw4w9WgXcQ

Output Formats

Output Directory

youtube-transcript/
├── .index.json                          # Video ID → directory path mapping (for cache lookup)
└── {channel-slug}/{title-full-slug}/
    ├── meta.json                        # Video metadata (title, channel, description, duration, chapters, etc.)
    ├── transcript-raw.json              # Raw transcript snippets from YouTube API (cached)
    ├── transcript-sentences.json        # Sentence-segmented transcript (split by punctuation, merged across snippets)
    ├── imgs/
    │   └── cover.jpg                    # Video thumbnail
    ├── transcript.md                    # Markdown transcript (generated from sentences)
    └── transcript.srt                   # SRT subtitle (generated from raw snippets, if --format srt)

• {channel-slug}: Channel name in kebab-case
• {title-full-slug}: Full video title in kebab-case

The --list mode outputs to stdout only (no file saved).

Caching

On first fetch, the script saves:

• meta.json — video metadata, chapters, cover image path, language info
• transcript-raw.json — raw transcript snippets from YouTube API ({ text, start, duration }[])
• transcript-sentences.json — sentence-segmented transcript ({ text, start: "HH:mm:ss", end: "HH:mm:ss" }[]), split by sentence-ending punctuation (.?!…。？！ etc.), timestamps proportionally allocated by character length, CJK-aware text merging
• imgs/cover.jpg — video thumbnail

Subsequent runs for the same video use cached data (no network calls). Use --refresh to force re-fetch. If a different language is requested, the cache is automatically refreshed.

When YouTube returns anti-bot / blocked responses on the direct InnerTube path, the script retries with alternate client identities and then falls back to yt-dlp if available. If fallback is needed but yt-dlp is unavailable, the agent should decide how to make yt-dlp available and continue rather than pushing the installation decision to the user.

SRT output (--format srt) is generated from transcript-raw.json. Text/markdown output uses transcript-sentences.json for natural sentence boundaries.

Workflow

When user provides a YouTube URL and wants the transcript:

1. Run with --list first if the user hasn't specified a language, to show available options
2. Always single-quote the URL when running the script — zsh treats ? as a glob wildcard, so an unquoted YouTube URL causes "no matches found": use 'https://www.youtube.com/watch?v=ID'
3. Default: run with --chapters --speakers for the richest output (chapters + speaker identification)
4. The script auto-saves cached data + output file and prints the file path
5. For --speakers mode: after the script saves the raw file, follow the speaker identification workflow below to post-process with speaker labels

When user only wants a cover image or metadata, running the script with any option will also cache meta.json and imgs/cover.jpg.

When re-formatting the same video (e.g., first text then SRT), the cached data is reused — no re-fetch needed.

Chapter & Speaker Workflow

Chapters (--chapters)

The script parses chapter timestamps from the video description (e.g., 0:00 Introduction), segments the transcript by chapter boundaries, groups snippets into readable paragraphs, and saves as .md with a Table of Contents. No further processing needed.

If no chapter timestamps exist in the description, the transcript is output as grouped paragraphs without chapter headings.

Speaker Identification (--speakers)

Speaker identification requires AI processing. The script outputs a raw .md file containing:

• YAML frontmatter with video metadata (title, channel, date, cover, description, language)
• Video description (for speaker name extraction)
• Chapter list from description (if available)
• Raw transcript in SRT format (pre-computed start/end timestamps, token-efficient)

After the script saves the raw file, spawn a sub-agent (use a cheaper model like Sonnet for cost efficiency) to process speaker identification:

1. Read the saved .md file
2. Read the prompt template at {baseDir}/prompts/speaker-transcript.md
3. Process the raw transcript following the prompt:
   • Identify speakers using video metadata (title → guest, channel → host, description → names)
   • Detect speaker turns from conversation flow, question-answer patterns, and contextual cues
   • Segment into chapters (use description chapters if available, else create from topic shifts)
   • Format with **Speaker Name:** labels, paragraph grouping (2-4 sentences), and [HH:MM:SS → HH:MM:SS] timestamps
4. Overwrite the .md file with the processed transcript (keep the YAML frontmatter)

When --speakers is used, --chapters is implied — the processed output always includes chapter segmentation.

Error Cases