Social Video Distill#

Distill short-form video without defaulting to local heavy lifting.

Prefer the fastest capable specialist first: existing captions, browser AI, then local fallback.

Workflow Decision Tree#

1. Clarify the output target.
   • Transcript only
   • Concise summary
   • Humor read / joke structure
   • Best line / key moment
   • Suggested caption or reply
2. Choose the lightest viable input path.
   • If the user already pasted a transcript: skip retrieval and distill directly.
   • If the platform likely exposes captions: use scripts/extract_captions.sh first.
   • If captions are missing but a browser AI session is available: use Gemini for distillation of any rough transcript, notes, or manually recovered dialogue.
   • If the task truly needs raw media and no better route exists: download media with scripts/download_media.sh and use a transcription fallback outside this skill.
3. Distill with browser AI before using local ASR when the goal is understanding rather than archival transcript quality.
   • Quick single-clip distillation: Gemini.
   • Multi-source/theme synthesis: NotebookLM.
4. Verify the result.
   • Separate direct transcript from inference.
   • If captions were unavailable and a rough transcript was used, say so.
   • Quote the strongest line exactly when possible.

Quick Start#

Install local helper runtimes once:

bash

bash skills/social-video-distill/scripts/install_runtime.sh

Try captions first:

bash

bash skills/social-video-distill/scripts/extract_captions.sh 'https://example.com/video'

Download media only when needed:

bash

bash skills/social-video-distill/scripts/download_media.sh 'https://example.com/video'

Ask Gemini to distill a prepared transcript or notes file:

bash

node skills/social-video-distill/scripts/ask_gemini_cdp.js \
  --prompt-file /absolute/path/to/prompt.txt

Preferred Operating Pattern#

1. Caption-first retrieval#

Use extract_captions.sh before downloading full media.

Good fit:

• YouTube videos with manual or auto captions
• platforms where yt-dlp can expose subtitles without full download

If captions exist:

• clean the text lightly
• keep timestamps only if they help the task
• send the cleaned transcript to Gemini or distill directly

2. Browser-first distillation#

Use Gemini when the job is:

• explain what this clip is saying
• summarize the point
• identify why it is funny
• extract the punchline
• draft a caption/reply

Use scripts/ask_gemini_cdp.js with a focused prompt file. Keep prompts short and task-specific.

If the user wants study-guide style synthesis across multiple clips or mixed sources, use NotebookLM instead of forcing Gemini into a long single-shot prompt.

3. Local fallback only when necessary#

Use local ASR only when:

• captions are unavailable
• browser AI cannot access the content
• the user explicitly wants a transcript and not just a distillation

Do not lead with local Whisper/faster-whisper just because it is available.

Output Guidance#

Default to a compact result with these fields when helpful:

• What it says
• What is directly supported
• What is inferred
• Best line
• Why it lands

For humor requests, prefer:

1. premise
2. escalation/twist
3. punchline

For social reply requests, give 2-3 options max.

When to Read References#

• Read references/prompts.md for ready-to-use distillation prompts.
• Read references/troubleshooting.md when captions, CDP, or Gemini interaction fails.