| name | g-skl-recon-yt |
| description | YouTube transcript ingestion into the vault. Uses yt-dlp to fetch transcripts locally — no Docker, no MCP, no screen captures. Stores in research/videos/ with analysis_depth=transcript_only for future vision upgrade. |
| token_budget | medium |
| subsystem_memberships | ["VAULT_AND_RESEARCH"] |
g-recon-yt
Scripts: .cursor/skills/g-skl-ingest-youtube/scripts/
Activate for: "ingest this YouTube video", "capture transcript", "add video to vault", "save YouTube notes", any YouTube URL.
Purpose
Capture YouTube video transcripts into the vault's research/videos/ folder using yt-dlp — a local command-line tool with no Docker or MCP requirement. Notes are stored as analysis_depth: transcript_only which signals a future task can upgrade them to full vision analysis when the MCP video analyzer is available.
Prerequisites
pip install yt-dlp
uv pip install yt-dlp
Verify: yt-dlp --version
Clean Room Boundary
This capture skill may store source material in the vault for research, provenance, and later review. When captured content is used for --deep, g-skl-res-deep, or any adoption workflow, downstream outputs must be clean-room reverse specs: behavior-level summaries, interfaces, workflows, data shapes, constraints, and architectural patterns only. Do not copy source code, comments, docs prose, prompts, tests, or unique strings into generated gald3r artifacts except tiny attributed excerpts that are license-compatible and necessary for review. Record source URL/path, license when discoverable, capture date, and extraction limits so reviewers can separate source evidence from original gald3r specifications.
Operation: INGEST_YOUTUBE
Capture a YouTube video transcript and create a vault note.
python .cursor/skills/g-skl-ingest-youtube/scripts/fetch_transcript.py \
--url "https://www.youtube.com/watch?v=VIDEO_ID" \
--vault-path /path/to/vault
What happens:
- Run
yt-dlp --dump-json to get metadata (title, channel, duration, upload date)
- Run
yt-dlp --write-auto-sub --skip-download --sub-format vtt to fetch captions
- Parse VTT → clean text (strip timestamps, deduplicate rolling captions)
- Generate 2–3 sentence summary and key-points list from transcript
- Write vault note to
research/videos/{YYYY-MM-DD}_{slugified-title}.md
- Append to
vault/log.md
- Add entry to
vault/_index.yaml
- Run
g-hk-vault-reindex.py (if available)
Output note path: {vault_root}/research/videos/{date}_{slug}.md
Operation: CHECK_SETUP
Verify yt-dlp is installed and working.
python .cursor/skills/g-skl-ingest-youtube/scripts/fetch_transcript.py --check-setup
Reports: yt-dlp version + a simple test that the binary is accessible.
Caption Source Priority
| Priority | Source | Notes |
|---|
| 1 | Manually uploaded captions (en) | Most accurate |
| 2 | Auto-generated captions (en) | Usually good for clear speech |
| 3 | Auto-translated to English | Fallback for non-English videos |
| 4 | No captions available | Metadata-only note with warning |
Vault Note Schema
---
date: 2026-04-06
type: video
ingestion_type: one_shot
source: https://www.youtube.com/watch?v=VIDEO_ID
title: "Video Title"
channel: "Channel Name"
published: "2026-01-15"
duration_minutes: 42
tags: [video]
refresh_policy: manual
source_volatility: snapshot
analysis_depth: transcript_only
source_notes: "Transcript captured 2026-04-06. No visual analysis."
---
{2–3 sentence AI-generated summary from transcript}
- {Point 1}
- {Point 2}
- {Point N}
{clean transcript text, timestamps removed}
Not In Scope
- No video download (no storage cost)
- No frame extraction or visual analysis (future upgrade via
analysis_depth field)
- No scheduled re-ingestion — these are one-shot captures
Obsidian Compatibility
All notes written by this skill must conform to VAULT_OBSIDIAN_STANDARD.md (§2 type registry, §3 tag taxonomy).
Required fields: type: video, tags: [video, ...], title:, date:.
Encoding: UTF-8 without BOM. See also: scripts/gen_vault_moc.py.
