youtube-extract — pull named artifacts from a YouTube video

You are the host agent running this skill. The user has asked for the resources, references, links, code, books, tools, or people mentioned in a video. Your job is to produce a categorized reference list at ~/youtube-reports/{date}-{slug}-{video_id}-extract.md.

This skill is purely factual reference extraction. It does not summarize, recommend, judge, or rank. Every entry is a verbatim mention from the transcript with a timestamp.

You make all LLM calls yourself using your own model and your existing auth — there is no Python orchestrator, no vendor SDK in this repo, and no API key required from the user. The only system requirement is Python 3.11+ with yt-dlp and youtube-transcript-api installed (Step 1.5 verifies this).

Workflow — follow these steps in order

Step 0 — Resolve skill paths

All scripts/… and prompts/… references in this document are relative to the directory containing this SKILL.md file — not the user's working directory. Before your first subprocess call, capture the absolute path to that directory (you already know it: it's the path you loaded this SKILL.md from). Use it as <SKILL_DIR> for every script and prompt path below.

In every shell call, pass quoted absolute paths:

python3 "<SKILL_DIR>/scripts/fetch.py" "<url>" --cache