Youtube Transcript Archiver

YouTube Transcript Archiver Skill

Batch-download and enrich transcripts from any YouTube channel or playlist into a structured markdown archive with extracted sources and downloadable assets.

What it does

1. Download phase: Fetches all episodes in a date range from a YouTube channel using yt-dlp. Produces one markdown file per episode containing title, date, duration, YouTube link, show notes, extracted links, and the full auto-generated transcript.

2. Enrichment phase: Scans each transcript for mentioned sources (reports, articles, papers, company announcements) using NLP pattern extraction. Cross-checks each extraction with Haiku for accuracy. Resolves URLs via Perplexity search where possible.

3. Asset phase: Downloads referenced reports/papers (PDFs, blog posts) to a local assets/ subdirectory and links them from each markdown file.

Usage

/youtube-archive <channel-url-or-handle> [--months N] [--output-dir path] [--enrich] [--assets]

Arguments

Examples

# Archive last 6 months of AI Daily Brief
/youtube-archive @AIDailyBrief --months 6 --enrich --assets

# Archive a specific playlist
/youtube-archive https://www.youtube.com/playlist?list=PLxxx --months 12

# Just transcripts, no enrichment
/youtube-archive @lexfridman --months 3 --output-dir lexfridman-transcripts

Prerequisites

• yt-dlp (installed via pip to ~/.local/lib/python3.12/site-packages)
• For enrichment: Haiku model access (via claude API or agent spawning)
• For asset downloads: Perplexity MCP tools (perplexity_search)
• For URL summarisation: WebFetch tool

Claude Code only: the enrichment/asset/swarm phases assume Claude Code affordances (Agent/Task spawning, WebFetch). On Codex / GPT-6 Astra: run the phases sequentially in one session and fetch URLs with a shell HTTP client instead of WebFetch.

Output structure

output-dir/
├── episode-title-slug.md          # One per episode
├── assets/
│   ├── mckinsey-ai-adoption-2026.pdf
│   └── vendor-technical-report-2026.pdf
├── source-index.md                # Cross-episode source index
└── .subs_tmp/                     # Temporary subtitle files (cleaned up)

Markdown file format

# Episode Title

- **Date**: 2026-08-18
- **Duration**: 29:13
- **YouTube**: https://www.youtube.com/watch?v=xxxxx

## Show Notes

(YouTube description)

## Links

- https://example.com/report.pdf → [local](assets/report.pdf)
- https://example.com/article

## Sources Mentioned

| Source | Type | Context | URL |
|--------|------|---------|-----|
| Bloomberg | article | "Bloomberg reports that..." | https://... |
| McKinsey AI report | report | "according to McKinsey's latest..." | [local](assets/...) |

## Transcript

(full transcript text, paragraphed every ~500 words)

Swarm deployment

For large archives (100+ episodes), deploy as a swarm:

Phase 1 (download):     1 agent, sequential (rate-limit friendly)
Phase 2 (enrichment):   N agents in parallel (one per batch of 10 episodes)
Phase 3 (asset fetch):  N agents in parallel (one per unique source)

The download phase is deliberately sequential to avoid YouTube rate limiting. Enrichment and asset phases parallelise well.

Implementation

The core download logic is in the yt-transcript-archive binary. It can be run standalone or invoked by agents.

Key design decisions

• --dump-json over --print: YouTube descriptions contain newlines that corrupt line-based metadata parsing. JSON output is reliable.
• SRT over VTT: Auto-generated VTT has overlapping cue timestamps that produce duplicate text. SRT conversion via yt-dlp deduplicates.
• Date-window filtering: The flat playlist API doesn't support date filtering, so we fetch all IDs (fast, metadata-only) then check each video's upload date individually. A streak of 15+ old episodes triggers early exit.
• Slug-based filenames: Episode titles are slugified to max 80 chars for filesystem safety. Video IDs are embedded in the markdown for deduplication.