Youtube To Vault

YouTube → Second Brain

Turn a YouTube video into a distilled, linked note in the Obsidian vault.

Vault location

$env:OBSIDIAN_VAULT — notes go in 03-Resources\. Follow CONVENTIONS.md and the _templates\video.md template in that vault.

Single video

1. Get the transcript (free, no install — uvx runs yt-dlp on demand):

``powershell $tmp = "$env:OBSIDIAN_VAULT\.tmp" New-Item -ItemType Directory -Force -Path $tmp | Out-Null uvx yt-dlp --skip-download --write-auto-subs --write-subs --sub-langs en --sub-format vtt --convert-subs vtt -o "$tmp\%(id)s.%(ext)s" "" `` Also grab metadata for title/creator: `powershell uvx yt-dlp --skip-download --print "%(title)s|%(uploader)s|%(id)s" "" `` If no subtitles exist, fall back to local Whisper transcription (next section) — do NOT give up.

Fallback: local Whisper (for videos with no captions)

The user has an RTX 5060 (8 GB). Transcribe the audio locally — free, no API. Uses whisper-ctranslate2 (a faster-whisper CLI) run on demand via uvx, GPU if available:

$tmp = "$env:OBSIDIAN_VAULT\.tmp"
New-Item -ItemType Directory -Force -Path $tmp | Out-Null
# 1) grab audio only (small + fast)
uvx yt-dlp -x --audio-format mp3 -o "$tmp\%(id)s.%(ext)s" ""
# 2) transcribe -> writes .vtt next to the audio
uvx whisper-ctranslate2 "$tmp\.mp3" --model small --language en `
  --output_format vtt --output_dir $tmp --device auto --compute_type auto

• --model small is the sweet spot for speed/accuracy on 8 GB; bump to medium for

hard audio, drop to base for speed. --device auto uses CUDA when present, else CPU.

• First run downloads the model weights once (cached).
• Then continue from step 2 below (read & distill the .vtt) exactly as normal.
• Flag to the user that Whisper output is a machine transcription — names/jargon may be off.

1. Read & distill the .vtt file. Strip timestamps/markup. Produce:

• a 1-paragraph Summary
• 5–10 Key takeaways, each prefixed with the [mm:ss] it came from
• any code/steps, in a code block

Keep it tight — distill, don't dump the raw transcript.

1. Write the note to 03-Resources\.md (slug = kebab-case title) using the

video.md template. Fill frontmatter: title, type: video, tags, source = the exact URL, creator = channel, created = today. Add a Questions this answers line. Add [[wikilinks]] to any related existing notes (search the vault first).

1. Flag that auto-transcribed code can be wrong — tell the user to verify snippets.

1. Delete the .tmp subtitle files when done.

Batch (a playlist / course) — use a parallel agent team

When given a playlist, list its videos first:

uvx yt-dlp --flat-playlist --print "%(url)s" ""

Then dispatch parallel agents, following the user's Agent Team Standard:

• Goal: one well-formed, cross-linked note per video.
• Workers (Sonnet 4.6 — good transcript distillation at reasonable cost; Haiku 4.5

for very short clips): each worker OWNS exactly one file 03-Resources\.md. Zero file overlap — partition by video.

• Lead (Opus 4.8 — synthesis): dedupe topics, add cross-[[links]] between the new

notes and existing vault notes, update index.md.

Notes & gotchas

• Needs a video with captions/auto-captions. Some have none.
• First uvx yt-dlp run downloads yt-dlp into a cache (one-time, a few seconds).
• Never save the raw transcript as the note — always distill.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: ridwanahmed06
• Source: ridwanahmed06/claude-skills
• License: MIT

Install and usage instructions live in the source repository linked above.