Skills
skills/mrpaulscrivens/boring-zoo/boring-youtube-mining

boring-youtube-mining

SKILL.md

YouTube Mining — World Code Edition

You pull recent videos from pre-selected YouTube channels, download their transcripts, and generate content ideas the user can actually use. One pass. No back-and-forth.

Before Starting — Check Dependencies

Run this silently:

which yt-dlp

If yt-dlp is NOT installed, stop and tell the user:

yt-dlp is required but not installed.

Install it for your platform:

  • macOS: brew install yt-dlp
  • Windows: winget install yt-dlp
  • Linux: pip install yt-dlp or your package manager
  • Universal: pip install yt-dlp

Then run /boring-youtube-mining again.

Do not proceed without yt-dlp. Stop gracefully.

Before Starting — Load World Code (Optional)

Try to read these files silently:

  • world-code/voice.md — Voice rules (NOT applied to output — this is research, not published content)
  • world-code/conversation.md — Bridge structure (walls, struggles, goblins, treasures)

If conversation.md exists: Use the Bridge to filter and prioritize ideas. Every idea gets mapped to a wall, struggle, or goblin. Ideas that don't connect to any get deprioritized (still listed, just flagged as "no Bridge connection").

If conversation.md doesn't exist: Generate ideas purely from transcript analysis. Don't nag about missing files.

Step 1: Load Channel List

Read settings/youtube-channels.md.

If the file doesn't exist, create it with starter content and tell the user:

I created settings/youtube-channels.md with some example channels. Edit it in Obsidian to add the channels you want to mine, then run /boring-youtube-mining again.

Starter content for settings/youtube-channels.md:

# YouTube Channels

Add channels you want to mine for content ideas. Use the @ handle format.
Add a short description so Claude knows the context.

- @AlexHormozi - Business, offers, scaling
- @ChrisDo - Branding, pricing, positioning

Stop after creating the file. Let the user configure it first.

Step 2: Fetch Recent Videos

For each channel in the list, run:

yt-dlp --flat-playlist --playlist-items 1:5 --print "%(id)s | %(title)s | %(duration_string)s" "https://www.youtube.com/@{handle}/videos" 2>/dev/null

Process channels sequentially (not in parallel) to avoid rate limiting.

Present the results as a numbered list:

## Recent Videos

### @AlexHormozi
1. [dQw4w9WgXcQ] How to Price Your Offer (12:34)
2. [abc123def] The $100M Framework Nobody Uses (18:22)
...

### @ChrisDo
3. [xyz789ghi] Why Your Brand Is Invisible (9:45)
...

Ask the user:

Which videos should I analyze? Enter numbers (e.g., "1, 3, 5") or "all".

If a channel handle fails to resolve, report it and continue with the remaining channels:

Could not fetch videos from @BadHandle — check the handle in settings/youtube-channels.md.

Step 3: Download Transcripts

For each selected video, first check if a cached transcript exists:

ls "content/youtube-transcripts/{video_id}.md" 2>/dev/null

If cached: Skip download, read from cache. Tell the user:

Transcript for "{title}" already cached — skipping download.

If not cached: Download the auto-generated English captions:

yt-dlp --write-auto-sub --sub-lang en --sub-format vtt --skip-download -o "content/youtube-transcripts/%(id)s" "https://www.youtube.com/watch?v={video_id}" 2>/dev/null

Then clean the VTT file and save as markdown. Create the directory if needed:

mkdir -p "content/youtube-transcripts"

Cleaning VTT Captions

The raw VTT file contains timestamps and duplicate lines. Clean it:

  1. Strip all timestamp lines (lines matching XX:XX:XX.XXX --> XX:XX:XX.XXX)
  2. Strip the WEBVTT header and Kind:/Language: metadata lines
  3. Remove duplicate consecutive lines (VTT repeats lines across cue boundaries)
  4. Remove position/alignment tags like <c>, </c>, align:start position:0%
  5. Join into paragraphs (group sentences, add line breaks at natural pauses)

Save as content/youtube-transcripts/{video_id}.md with frontmatter:

---
video_id: {video_id}
title: {video title}
channel: {channel handle}
date_fetched: {YYYY-MM-DD}
url: https://www.youtube.com/watch?v={video_id}
---

# {video title}

**Channel:** {channel handle}
**URL:** https://www.youtube.com/watch?v={video_id}

## Transcript

{cleaned transcript text}

Delete the raw VTT file after conversion.

Edge Cases

  • No English auto-captions available: Skip the video with a message:

    No English captions available for "{title}" — skipping.

  • Very long transcripts (estimated 10,000+ words): Chunk-summarize before idea generation. Break into ~3,000 word sections, summarize each section's key points, then use the summaries for idea generation. Note in the output that the transcript was summarized.

Step 4: Generate Ideas

For each transcript, generate ideas through two lenses:

Lens 1: Direct Response

How would the user respond to this video's ideas? Generate 2-3 ideas:

  • Agree & Expand — They made a point you also believe. What's your unique angle on it?
  • Disagree & Counter — They said something you see differently. What's your take?
  • Build On — They touched on something but stopped short. Where would you take it?

Lens 2: Gap Ideas

What's missing? Generate 1-2 ideas:

  • Topics they mentioned but didn't go deep on
  • Adjacent topics their audience would care about but weren't covered
  • The question their video raises but doesn't answer

Per Idea, Include:

Field Description
Idea title Sharp, specific — not "Thoughts on pricing" but "Why Hourly Pricing Kills Solo Businesses"
Angle Direct Response (agree/disagree/expand) or Gap
Source Video title + timestamp range if identifiable
Your take (1-2 sentences) The core argument you'd make
Bridge mapping Which wall/struggle/goblin this connects to (if conversation.md exists)
Content type Post, thread, email, essay, video script, carousel
Draft hook One opening line that would stop the scroll

Quality Over Quantity

  • 3-5 ideas per video. Not 10. Not 20. The best 3-5.
  • Every idea must pass the "would I actually make this?" test
  • If an idea is generic ("Content is important"), kill it
  • Ideas without a clear angle aren't ideas — they're topics. Topics aren't useful.

Bridge-First Filtering (when conversation.md exists)

After generating all ideas, sort them:

  1. Strong Bridge connection — directly maps to a wall, struggle, or goblin
  2. Loose Bridge connection — related to the user's world but not a direct map
  3. No Bridge connection — interesting but disconnected from their World Code

Group 3 still gets listed but flagged. The user decides if it's worth pursuing.

Step 5: Save Output

Write the output file to content-ideas/youtube-mining-{YYYY-MM-DD}.md:

---
type: youtube-mining
date: {YYYY-MM-DD}
channels_mined: [{list of channels}]
videos_analyzed: {count}
ideas_generated: {count}
---

# YouTube Mining — {YYYY-MM-DD}

## Sources

| Video | Channel | Ideas |
|-------|---------|-------|
| {title} | {channel} | {count} |
...

## Ideas

### From: "{video title}" (@channel)

#### 1. {Idea Title}

- **Angle:** {Direct Response — Agree/Disagree/Expand | Gap}
- **Source:** {video title}, ~{timestamp context if available}
- **Your take:** {1-2 sentences}
- **Bridge:** {wall/struggle/goblin or "No direct connection"}
- **Content type:** {post/thread/email/essay/video/carousel}
- **Draft hook:** "{opening line}"

---

{repeat for each idea}

## Bridge Summary

### Strong Connections
- {idea} → {wall/struggle}
...

### No Connection
- {idea} — interesting but not tied to your current World Code
...

## Next Steps

- Pick an idea and run `/boring-copywriting` to draft it
- Run `/boring-social-content` to turn an idea into platform-specific posts
- Run `/boring-remix` to generate multiple angles on your favorite idea

Step 6: Wrap Up

After saving, tell the user:

Saved {X} ideas to content-ideas/youtube-mining-{date}.md.

{count} ideas with strong Bridge connections, {count} without.

Pick an idea and I can draft it with /boring-copywriting, turn it into social posts with /boring-social-content, or remix it with /boring-remix.

That's it. No recap of the process. No lecture on content strategy.

Key Principles

  • One pass, full output. Fetch → transcripts → ideas → save. No stopping to ask "should I continue?"
  • Research output, not published content. No voice applied. No polishing. Raw strategic ideas.
  • Caching is non-negotiable. Never re-download a transcript. Check the cache first, always.
  • Bridge-first when available. The World Code connection is what makes this useful instead of just interesting.
  • Sequential channel processing. Don't hammer YouTube with parallel requests.
  • 3-5 quality ideas per video. Restraint is the skill. Anyone can generate 50 mediocre ideas.
  • Draft hooks matter. An idea without a hook is just a topic. Topics are cheap.
  • Don't explain the process. The user doesn't need a play-by-play of "now I'm downloading transcripts." Just do it and show the results.

Related Skills

  • boring-copywriting — Draft full content from a mined idea
  • boring-social-content — Turn ideas into platform-specific posts
  • boring-remix — Generate multiple angles on a single idea
  • boring-content-strategy — Broader content planning using mined insights
Weekly Installs
17
Repository
mrpaulscrivens/…ring-zoo
GitHub Stars
3
First Seen
8 days ago
Security Audits
Gen Agent Trust HubPass
SocketFail
SnykWarn
Installed on
opencode17
github-copilot17
codex17
amp17
cline17
kimi-cli17