g-skl-recon-yt

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
# or
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:

1. Run yt-dlp --dump-json to get metadata (title, channel, duration, upload date)
2. Run yt-dlp --write-auto-sub --skip-download --sub-format vtt to fetch captions
3. Parse VTT → clean text (strip timestamps, deduplicate rolling captions)
4. Generate 2–3 sentence summary and key-points list from transcript
5. Write vault note to research/videos/{YYYY-MM-DD}_{slugified-title}.md
6. Append to vault/log.md
7. Add entry to vault/_index.yaml
8. 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."
---

## Summary
{2–3 sentence AI-generated summary from transcript}

## Key Points
- {Point 1}
- {Point 2}
- {Point N}

## Full Transcript
{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.