YouTube

Fetch the full transcript of any YouTube video and turn it into whatever the user needs — summaries, blog posts, social content, quotes, show notes, or raw text. Then package videos for upload: high-CTR thumbnails, SEO titles, and descriptions.

Routing

Requirements

• Hyper MCP installed. https://app.hyperfx.ai/mcp
• Sandbox text workflows: require ai_functions_run in the connected catalog and the sandbox toolkit. These scripts run inline LLM calls through the sandbox tool bridge.
• YouTube toolkit enabled at https://app.hyperfx.ai/apps — provides youtube_video_transcripts_fetch and youtube_videos_read.
• Thumbnail workflows additionally need the image generation and sandbox toolkits.

If youtube_video_transcripts_fetch is not found, the toolkit to enable is YouTube (youtube_toolkit).

How to run the tools in this skill

Every tool in this skill is named by its canonical tool name. Run it with the call your surface gives you:

If a tool is not found: in Hyper chat, find its toolkit id with discover_toolkits and pass it to enable_toolkit. On an MCP client or the CLI, tell the user which integration to connect in Hyper.

Two tools — pick the right one

Default: start with youtube_video_transcripts_fetch. Use youtube_videos_read when you need something the raw text can't give you (e.g. visual descriptions, translation, or a structured extraction from a very long video).

Critical rules

1. youtube_video_transcripts_fetch takes 15–30 seconds. It spins up an isolated sandbox. Tell the user it's running and to expect a short wait — don't make them think it's stuck.
2. Both video IDs and full URLs are accepted. "NZLAdOL9fP8" and "https://www.youtube.com/watch?v=NZLAdOL9fP8" both work.
3. Don't fabricate transcript content. Always fetch before summarizing. Never rely on training knowledge about what a specific video says.
4. Very long videos (>2 hours): youtube_video_transcripts_fetch handles these fine. Only use youtube_videos_read on long videos if you specifically need AI-powered extraction — it can hit token limits on very long content.
5. No transcript available: Some videos have transcripts disabled. If youtube_video_transcripts_fetch fails, try youtube_videos_read as a fallback — it uses a different extraction method.

Fetching the transcript

youtube_video_transcripts_fetch(
    video_id_or_url="https://www.youtube.com/watch?v=NZLAdOL9fP8",
    language="en"   # optional — omit to auto-detect
)

Response structure:

{
  "success": true,
  "video_id": "NZLAdOL9fP8",
  "language": "English (auto-generated)",
  "text": "Full transcript as one string...",
  "segments": [
    { "text": "This week we launched Hyper MCP.", "start": 0.0, "duration": 3.2 },
    { "text": "It brings Hyper's built-in tools...", "start": 3.2, "duration": 4.1 }
  ],
  "total_duration": 342.0
}

Use text for most tasks. Use segments when you need timestamps (e.g. chapters, clip references, karaoke captions).

Using youtube_videos_read for AI-powered extraction

youtube_videos_read(
    url="https://www.youtube.com/watch?v=NZLAdOL9fP8",
    instruction="Summarize the key points. Then list the main features demonstrated, with timestamps."
)

Good instruction examples:

• "Extract every claim made about pricing or cost."
• "List the action items mentioned, in order."
• "Translate this to Spanish."
• "What tools or products does the speaker mention by name?"
• "Identify the main sections of this video and give me a timestamp for each."

What to do with the transcript

Once you have the text, ask the user what they need — or infer it from context:

Thumbnails and SEO packaging

For making or refreshing thumbnails, cloning the style of top-ranking thumbnails, adding the user's face, and generating SEO titles/descriptions, read references/thumbnails.md. Every thumbnail workflow is a sandbox script under scripts/ (generate_thumbnail.py, research_top_thumbnails.py, clone_top_thumbnail_style.py) — the reference doc is the routing table and the rules for using them.

Example outputs

Input: "Get the transcript of https://www.youtube.com/watch?v=NZLAdOL9fP8 and write a LinkedIn post from it"

Flow:

1. Call youtube_video_transcripts_fetch(video_id_or_url="https://www.youtube.com/watch?v=NZLAdOL9fP8")
2. Read the returned text
3. Identify the 1–2 sharpest moments — what's surprising, useful, or quotable
4. Draft a LinkedIn post in the speaker's voice (first person) with a hook and a clear point

Input: "Summarize this video for me: [URL]"

Flow:

1. Call youtube_video_transcripts_fetch(video_id_or_url="[URL]")
2. Return 4–6 bullet points of key takeaways, without padding or filler

Input: "Make me a thumbnail like the top videos for 'AI agents'"

Flow:

1. Read references/thumbnails.md
2. Run scripts/clone_top_thumbnail_style.py with query="AI agents" and the user's topic
3. Show the top thumbnails, let the user pick a rank, re-run with chosen_rank to generate

Related skills

For title options, descriptions and thumbnail concepts, run generate_seo_titles.py, generate_seo_description.py and analyze_thumbnail_concepts.py under scripts/. They call ai_functions_run from the sandbox using fetched transcript and video context. Follow the output contracts in references/packaging-schemas.json; the thumbnail reference explains the workflow.