yt2skill

name	yt2skill
description	Turn a YouTube video into a reusable agent skill (or subagent). Fetches the transcript, extracts the teachable procedure, and writes an installable SKILL.md. Use when the user gives a YouTube URL and wants its method codified — 'make a skill from this video', 'yt2skill <url>'.
argument-hint	<youtube-url> [--agent] [--name <slug>]
allowed-tools	["Bash","Read","Write","AskUserQuestion"]

Input: a YouTube URL. Output: a new, installed agent skill (or subagent) that captures the *transferable method* taught in the video — not a summary of the video. The video is the source; the skill is a procedure the agent can re-run.

Flow

1. Get the transcript. From this skill's own directory:

D="$(dirname "$(readlink -f ~/.claude/skills/yt2skill/SKILL.md)")"
python3 "$D/fetch_transcript.py" "<URL>" > /tmp/yt2skill.txt; echo "exit=$?"

Handle the exit code:

• 0 — read /tmp/yt2skill.txt.
• 2 (NEED_INSTALL) — python3 -m pip install --user --quiet youtube-transcript-api, then rerun.
• 3 (no captions) — tell the user this video has no transcript; offer to let them paste one, or stop. Do not invent content.
• 4 (bad url) — ask for a valid URL.

2. Read the transcript and extract the method. Identify what the video teaches you to do: the steps, rules, heuristics, gotchas, and order of operations. Strip filler, sponsorships, and intros. If the video is pure commentary with no transferable procedure, say so and stop — not everything is a skill.

3. Decide skill vs subagent (default: skill):

• Skill — a procedure/checklist the agent runs on request. The common case.
• Subagent (--agent, or when the video describes a role/persona doing open-ended judgement work) — write a ~/.claude/agents/<slug>.md instead, with name/description/tools frontmatter and a system prompt in the persona.

4. Write it. Pick a short kebab-case <slug> (or use --name). For a skill:

~/.claude/skills/<slug>/SKILL.md:

---
name: <slug>
description: "<one line: what it does + when to use it, third person>"
---

# <Title>

<2-3 lines: what method this is and where it came from (video URL).>

## Steps
1. ...
2. ...

## Rules / gotchas
- ...

Rules for a good generated skill:

• Imperative steps the agent can act on, not prose about the video.
• Concrete: thresholds, commands, orderings — whatever the video specified.
• Keep the video URL in a one-line Source: footer for provenance.
• Ponytail: only the steps that matter. No invented sections, no boilerplate.

5. Confirm + report. Show the path and the one-line description. The skill is live immediately (~/.claude/skills/ is auto-discovered); a subagent likewise. Tell the user the trigger (/<slug> or the agent name).

Notes

• No captions → no skill. Whisper-from-audio fallback is intentionally not built (heavy: yt-dlp + ffmpeg + model). Add it only if a user hits this repeatedly.
• Don't overfit to one example in the video — capture the general method.