yt-dlp-archiver

yt-dlp Video Archiver

Archive YouTube videos with full metadata extraction, then analyze the content with fabric skills. Requires yt-dlp to be installed.

Preflight

Before doing anything, verify yt-dlp exists:

which yt-dlp || { echo "ERROR: yt-dlp is not installed. Install it first: brew install yt-dlp"; exit 1; }

If yt-dlp is not found, stop and tell the user. Do not attempt to install it.

Archive Directory Structure

All downloads go to ~/src/youtube-archive/ organized by date and video title:

~/src/youtube-archive/
  YYYYMMDD/
    YYYYMMDD-<sanitized-title>.mp4              # 480p video (sponsor segments removed)
    YYYYMMDD-<sanitized-title>.en.srt           # English subtitles (timecoded SRT)
    YYYYMMDD-<sanitized-title>.transcript.txt   # Clean transcript (no timestamps)
    YYYYMMDD-<sanitized-title>.description      # Video description
    YYYYMMDD-<sanitized-title>.info.json        # Full metadata JSON
    YYYYMMDD-<sanitized-title>.sponsorblock.json # SponsorBlock segment data
    YYYYMMDD-<sanitized-title>.jpg              # Thumbnail
    YYYYMMDD-<sanitized-title>.fabric.md        # Fabric analysis output (auto-generated)
    chapters/                                    # Chapter frame captures (if chapters exist)
      ch01-<chapter-title>.jpg
      ch02-<chapter-title>.jpg

The date folder uses today's date (the archive date), not the upload date.

The Download Command

Run this as a single yt-dlp invocation. yt-dlp handles everything in one pass — do not run multiple commands for the same video.

DATE=$(date +%Y%m%d)
mkdir -p ~/src/youtube-archive/$DATE

yt-dlp \
  -f "bestvideo[height<=480]+bestaudio/best[height<=480]" \
  --merge-output-format mp4 \
  --restrict-filenames \
  --write-auto-subs \
  --sub-lang "en" \
  --convert-subs srt \
  --write-description \
  --write-info-json \
  --write-thumbnail \
  --convert-thumbnails jpg \
  --sponsorblock-remove all \
  --embed-chapters \
  --no-playlist \
  --output "~/src/youtube-archive/$DATE/${DATE}-%(title)s.%(ext)s" \
  "<URL>"

What each flag does

• -f "bestvideo[height<=480]+bestaudio/best[height<=480]" — caps video at 480p, grabs best audio
• --merge-output-format mp4 — ensures final output is mp4 regardless of source streams
• --restrict-filenames — forces ASCII-safe filenames (replaces spaces with _, strips unicode)
• --write-auto-subs --sub-lang en --convert-subs srt — downloads auto-generated English captions as .srt
• --write-description — saves the video description as a .description text file
• --write-info-json — dumps the full metadata blob (title, channel, views, likes, duration, chapters, categories, tags, etc.)
• --write-thumbnail --convert-thumbnails jpg — grabs the thumbnail and converts to jpg
• --sponsorblock-remove all — removes sponsor segments, intros, outros, selfpromo from the actual video file
• --embed-chapters — bakes chapter markers into the mp4
• --no-playlist — if the URL is part of a playlist, only download this single video

After Download

Once the download completes:

1. Report what was downloaded — list the files with sizes
2. Show a brief summary extracted from the info.json: title, channel, duration, upload date, view count, chapter count
3. If chapters exist, list them with timestamps
4. Run the Post-Processing Pipeline (next section)
5. Then proceed to the Fabric Analysis Pipeline

Post-Processing Pipeline

Run these three steps after yt-dlp finishes but before fabric analysis.

1. Generate clean transcript

Strip SRT timestamps to produce a plain text transcript. The SRT format has: line number, timestamp line, subtitle text, blank line. Extract only the text:

sed '/^[0-9][0-9]*$/d; /^[0-9][0-9]:[0-9][0-9]:[0-9][0-9]/d; /^$/d' \
  "YYYYMMDD-<title>.en.srt" > "YYYYMMDD-<title>.transcript.txt"

If no .en.srt exists, skip this step. The fabric pipeline will use the .description file instead.

Use the .transcript.txt (not the .en.srt) as input for the fabric analysis pipeline. The clean transcript produces better analysis than timecoded subtitles.

2. Fetch SponsorBlock segment data

Query the SponsorBlock API to get segment data for the video. Extract the video ID from the URL (the v= parameter or the path after youtu.be/):

VIDEO_ID="<extracted-video-id>"
curl -s "https://sponsor.ajay.app/api/skipSegments?videoID=${VIDEO_ID}&categories=%5B%22sponsor%22%2C%22selfpromo%22%2C%22interaction%22%2C%22intro%22%2C%22outro%22%2C%22preview%22%2C%22music_offtopic%22%2C%22filler%22%5D" \
  -o "YYYYMMDD-<title>.sponsorblock.json"

If the API returns a 404 (no segments found for this video), write an empty JSON array [] to the file. Report a summary of what was found: segment count, total removed duration, and segment types.

3. Extract chapter frame captures

If the info.json contains a chapters array, extract a single frame from the video at each chapter's start time. Requires ffmpeg to be installed:

ffmpeg -ss <start_time_seconds> -i "YYYYMMDD-<title>.mp4" \
  -frames:v 1 -q:v 2 \
  "chapters/ch<NN>-<sanitized-chapter-title>.jpg"

Run this for each chapter. Create the chapters/ subdirectory first. Sanitize chapter titles the same way yt-dlp does (ASCII-safe, underscores for spaces).

If ffmpeg is not installed, skip this step and note it in the download report. If no chapters exist, skip this step entirely.

Fabric Analysis Pipeline

After downloading, chain fabric skills from the-no-shop marketplace against the transcript to produce a .fabric.md file. Each fabric skill is a standalone SKILL.md containing its own identity, steps, and output instructions — read each one and follow its instructions against the transcript content.

How chaining works

1. Read the .transcript.txt (or .en.srt if no txt, or .description if no transcript at all)
2. For each skill in the pipeline, find its SKILL.md using Glob with path set to the marketplace directory. Try these locations (first match wins):
   • Installed: $HOME/.claude/plugins/marketplaces/the-no-shop/plugins/ with pattern fabric-*/skills/<skill-name>/SKILL.md
   • Local dev: working directory plugins/ with same pattern
3. Read the skill's SKILL.md
4. Follow its IDENTITY/PURPOSE, STEPS, and OUTPUT INSTRUCTIONS sections against the transcript content
5. Write that skill's output as a section in the .fabric.md file
6. Move to the next skill in the chain

When applying each fabric skill, temporarily adopt its stated identity and follow its instructions faithfully. Each skill's output is independent — do not let one skill's framing influence the next.

Core pipeline (always chain these)

Run these two fabric skills in order against every transcript:

Order	Skill	Plugin	What it produces
1	summarize	fabric-summarization	ONE SENTENCE SUMMARY, MAIN POINTS, TAKEAWAYS
2	extract-wisdom	fabric-extraction	IDEAS, INSIGHTS, QUOTES, HABITS, FACTS, REFERENCES, ONE-SENTENCE TAKEAWAY, RECOMMENDATIONS

Content-aware chain additions

After reading the info.json, detect the content type and chain one additional fabric skill from the analysis plugin:

Content type	Detection signal	Skill (plugin: fabric-analysis)
Debate/politics	title/description contains "debate"; categories include "News & Politics"	analyze-debate
Presentation/talk	title contains "talk", "keynote", "presentation", "conference"	analyze-presentation
Academic/paper	categories include "Education", "Science & Technology"; lecture-style content	analyze-paper

If the content doesn't match any of these types, skip the additional skill — the core pipeline (summarize + extract-wisdom) is sufficient.

Output file

Write all chained skill outputs to YYYYMMDD-<title>.fabric.md alongside the other downloaded files.

File header (always include)

# Fabric Analysis: <video title>

**Source**: <youtube URL>
**Channel**: <channel name>
**Duration**: <duration>
**Archived**: <today's date>
**Skills chained**: summarize, extract-wisdom[, additional-skill-name]

## Archive Manifest

| File | Size | Description |
|------|------|-------------|
| `YYYYMMDD-<title>.mp4` | 172MB | 480p video (sponsors removed) |
| `YYYYMMDD-<title>.transcript.txt` | 151K | Clean transcript (no timestamps) |
| `YYYYMMDD-<title>.fabric.md` | 21K | Fabric analysis (this file) |
| `YYYYMMDD-<title>.info.json` | 1.0M | Full metadata |
| `YYYYMMDD-<title>.en.srt` | 249K | Timecoded subtitles |
| `YYYYMMDD-<title>.jpg` | 68K | Thumbnail |
| `YYYYMMDD-<title>.description` | 2.9K | Video description |
| `YYYYMMDD-<title>.sponsorblock.json` | 841B | SponsorBlock segments |
| `chapters/` | 9 files | Chapter frame captures |

Populate the table from actual file sizes on disk (use ls -lh output). The Description column should reflect what was actually produced:

• For .mp4: include resolution and whether sponsors were removed
• For .transcript.txt: "Clean transcript (no timestamps)"
• For .fabric.md: list the chained skills, e.g. "summarize + extract-wisdom + analyze-paper"
• For .sponsorblock.json: include segment count and total removed duration if available
• For chapters/: include file count
• Omit rows for files that don't exist (e.g. no chapters/ if no chapters)

Then write each skill's output in order, using the section headers that each skill's instructions specify.

Formatting overrides

Each fabric skill has its own output formatting rules. Follow them, with these overrides for consistency across the combined document:

• Use markdown headers (##) to separate each skill's output sections
• Do not repeat items across skills — if extract-wisdom produces quotes that summarize already covered, skip the duplicates
• Do not output meta-commentary about the chaining process — only the analysis

If a fabric skill is not found

If the Glob search doesn't locate a skill's SKILL.md (the plugin isn't installed), skip it and add a note at the bottom of the .fabric.md: <!-- Skill not found: <skill-name> — skipped -->

Music videos and copyrighted lyrics

When the video categories include "Music" or the content is primarily a music video/song, the fabric pipeline will likely fail because the API content filter blocks reproduction of copyrighted lyrics. In this case:

1. Skip transcript-based fabric analysis entirely
2. Run fabric skills on the .description file only
3. Note in the .fabric.md that analysis is description-only due to music content
4. Do not attempt to quote, summarize, or paraphrase song lyrics

This is an API-level constraint, not a bug — the content filter prevents copyrighted lyric reproduction in model output.

When transcript is unavailable

If no .srt was generated (video has no captions), run fabric skills on the .description file instead. Note in the output that analysis is description-only.

Audio-Only Mode

If the user says "just the audio", "mp3 only", "rip the audio", or similar — replace the video download flags with:

yt-dlp \
  -x --audio-format mp3 \
  --restrict-filenames \
  --write-auto-subs \
  --sub-lang "en" \
  --convert-subs srt \
  --write-description \
  --write-info-json \
  --write-thumbnail \
  --convert-thumbnails jpg \
  --no-playlist \
  --output "~/src/youtube-archive/$DATE/${DATE}-%(title)s.%(ext)s" \
  "<URL>"

Skip --sponsorblock-remove and --embed-chapters for audio-only (not applicable to mp3). Still run the fabric pipeline on the transcript.

Handling Multiple URLs

If the user pastes multiple YouTube URLs in one message, process each one sequentially. Each gets its own download in the same date folder.

Error Handling

• Age-restricted videos: Some may require authentication. Tell the user and suggest --cookies-from-browser chrome
• No auto-captions available: Some videos have no captions at all. The download still succeeds — just note that no transcript was generated and fabric will use description-only
• Geo-blocked: Report the error, suggest VPN or --geo-verification-proxy
• Live streams: These won't have a fixed duration. Note this to the user and suggest --live-from-start if they want to capture it

What NOT to do

• Do not attempt to install yt-dlp
• Do not download above 480p unless the user explicitly asks for higher quality
• Do not download entire playlists unless explicitly asked
• Do not run --write-comments by default (slow, API-heavy, usually not wanted)
• Do not skip the fabric pipeline — it runs automatically after every download