video-summarizer（视频摘要器）

Overview

This skill transcribes and summarizes video content by running the python -m summarizer CLI tool. It handles YouTube (captions or audio), local files, Google Drive, and Dropbox. Transcripts are chunked, processed in parallel through an OpenAI-compatible LLM, and merged into a final summary.

IMPORTANT: Always use the CLI command below. Never try to fetch, scrape, or download video URLs directly (e.g., with webfetch or curl). The CLI handles all downloading, transcription, and summarization internally.

Quick Start (Step-by-Step)

Follow these steps exactly:

Step 1 -- Run the CLI:

python -m summarizer --source "VIDEO_URL"

The tool uses the default provider from summarizer.yaml (already configured). No extra flags needed for basic usage.

Step 2 -- Read the output file:

The CLI prints the output filename on success. It is always saved inside the summaries/ subdirectory relative to the project working directory. For example:

[+] Saved watch_20260207_234533.md

The full path to read is: summaries/watch_20260207_234533.md

Step 3 -- Show the result to the user.

That's it. Three steps.

Where Files Live

All paths are relative to the project working directory (where summarizer.yaml and setup.py are).

Do NOT look for .env or config files in your home directory or in the skill directory. They are in the project root.

CLI Reference

python -m summarizer [OPTIONS]

Required

Provider Options

Source Options

Processing Options

Output Options

Warning: Do NOT use --no-save on Windows. There is a known encoding bug where Unicode characters in the summary crash stdout output. Always let the tool save to a file, then read the file.

Config Options

Summary Styles

Use with --prompt-type. Defined in summarizer/prompts.json.

Examples

Summarize a YouTube video (simplest form)

python -m summarizer --source "https://youtube.com/watch?v=VIDEO_ID"
# Output: summaries/watch_YYYYMMDD_HHMMSS.md

Specify a provider

python -m summarizer --source "https://youtube.com/watch?v=VIDEO_ID" --provider gemini
# Output: summaries/watch_YYYYMMDD_HHMMSS.md

Extract key insights with Distill Wisdom

python -m summarizer \
  --source "https://youtube.com/watch?v=VIDEO_ID" \
  --provider openrouter \
  --prompt-type "Distill Wisdom"
# Output: summaries/watch_YYYYMMDD_HHMMSS.md

Fact-check video claims using Perplexity

python -m summarizer \
  --source "https://youtube.com/watch?v=VIDEO_ID" \
  --base-url "https://api.perplexity.ai" \
  --model "sonar-pro" \
  --prompt-type "Fact Checker"
# Output: summaries/watch_YYYYMMDD_HHMMSS.md

Generate a Mermaid diagram from a lecture

python -m summarizer \
  --source "https://youtube.com/watch?v=VIDEO_ID" \
  --provider gemini \
  --prompt-type "Mermaid Diagram"
# Output: summaries/watch_YYYYMMDD_HHMMSS.md

Summarize a local file

python -m summarizer \
  --type "Local File" \
  --source "./lecture.mp4" \
  --provider groq
# Output: summaries/lecture_YYYYMMDD_HHMMSS.md

Batch-process multiple videos

python -m summarizer \
  --source "URL1" "URL2" "URL3" \
  --provider gemini \
  --prompt-type "Summarization"
# Output: one file per URL in summaries/

Use without config file (all flags explicit)

python -m summarizer \
  --source "https://youtube.com/watch?v=VIDEO_ID" \
  --base-url "https://openrouter.ai/api/v1" \
  --model "google/gemini-2.0-flash-exp:free" \
  --api-key "sk-or-v1-YOUR_KEY" \
  --prompt-type "Tutorial" \
  --no-config
# Output: summaries/watch_YYYYMMDD_HHMMSS.md

Reading the Output

After running the CLI, always read the output file from summaries/:

# The CLI prints something like:
# [+] Saved watch_20260207_234533.md
#
# The file is at:
summaries/watch_20260207_234533.md

Use the Read tool on the full path: summaries/<filename> (relative to project root).

Multi-Step Workflow

For comprehensive analysis, chain multiple styles on the same video:

1. Start with Summarization for a quick overview
2. Use Distill Wisdom to extract key insights
3. Run Fact Checker (ideally with Perplexity) to verify claims
4. Generate a Mermaid Diagram for visual reference

Error Handling

• If no API key is found, the tool checks .env for a key matching the provider URL keyword
• If YouTube captions are unavailable, the tool falls back to audio download + Whisper transcription
• --language auto does not force English. It uses the first available YouTube caption track and lets Whisper detect language.
• Use --verbose to see detailed progress and debug issues
• Cloud Whisper requires a Groq API key (free tier available)

Configuration Files

All in the project working directory:

• summarizer.yaml: Provider definitions (base_url, model, chunk-size) and defaults
• .env: API keys, auto-matched by URL keyword (e.g., generativelanguage = YOUR_KEY)
• summarizer/prompts.json: Prompt templates using {text} as transcript placeholder

Testing

python -m pytest tests/ -v