youtube-transcript

---

name: youtube-transcript disable-model-invocation: true description: This skill should be used when the user wants to extract a transcript from a YouTube video, summarize a YouTube video, or critically analyze a YouTube video's claims. Trigger on keywords like "youtube", "transcript", "summarize video", "critique video", or when a YouTube URL is provided. allowed-tools: Read, Write, Glob, Grep, Task, Bash(cat:), Bash(ls:), Bash(yt-dlp:), Bash(deno:), Bash(firefox:), Bash(mkdir:) argument-hint: [critique|summarize]

YouTube Transcript & Summary

Input: $ARGUMENTS

Prerequisites

• yt-dlp - transcript extraction
• firefox - cookie auth, opening results

Parse Arguments

• First word is critique → critique mode, URL is the rest
• First word is summarize → summarize mode, URL is the rest
• Otherwise → transcript-only mode, entire input is the URL

Step 1: Download subtitles + metadata (single yt-dlp call)

Combine subtitle download and metadata extraction into one call. Use a fixed output name to avoid special characters in video titles breaking shell commands:

mkdir -p /tmp/yt-transcript
yt-dlp --cookies-from-browser=firefox --write-auto-sub --write-sub --sub-lang en --skip-download --force-overwrites --output "/tmp/yt-transcript/video.%(ext)s" --print "%(title)s\n%(channel)s\n%(description)s\n%(duration_string)s" "VIDEO_URL"

If --cookies-from-browser=firefox fails, try chrome, then no cookies flag.

In parallel (if summarize/critique mode), also ensure the output directory exists:

mkdir -p "$(pwd)/summaries"

Step 2: Convert VTT to plain text

Auto-generated VTT files are extremely verbose (timestamps, positioning, HTML tags, duplicate overlapping lines). Strip them to plain text:

deno run --allow-read --allow-write $CLAUDE_SKILL_DIR/vtt-to-text.ts /tmp/yt-transcript/video.en.vtt /tmp/yt-transcript/transcript.txt

Read the resulting .txt file.

If mode is transcript-only: report file path, language, line count. Stop here.

Step 3: Delegate writing to Sonnet subagent

Use the Task tool with model: sonnet and subagent_type: general-purpose. Pass everything inline in the prompt so the subagent needs only a single Write tool call:

• The full transcript text (paste it into the prompt)
• The video metadata (title, channel, duration)
• The template (copy from below)
• The computed output file path
• Today's date

The subagent prompt should contain all data and end with: "Write the summary to {path} using the Write tool. Do not read any files - everything you need is above."

Output path

./summaries/{YYYY-MM-DD}_{slug}.md inside the current working directory (slug: lowercase kebab-case from title, max 60 chars). Always write into the repo the skill was invoked from.

Summarize template

---
created: {date}
type: summary
source: {url}
speaker: {name}
event: {channel}
---

## Overview

2-3 sentences: video length, speaker/channel, core topic/thesis.

## Key Findings

- Concise bullet points capturing the most important takeaways
- Focus on actionable insights and novel information
- 5-15 bullets depending on video length

Critique template

Read $CLAUDE_SKILL_DIR/critique.md for the full critique template with examples.

Step 4: Cleanup, report, and open

1. Clean up: rm /tmp/yt-transcript/transcript.txt
2. Print key findings in chat - show the Overview and Key Findings sections directly in the conversation so the user doesn't have to open the file to see the result.
3. Open result in Firefox: firefox "$path" &