Summarize Skill: One Command to Distill Any URL, PDF, or YouTube Video

30,000+ downloads and 143 stars — summarize by @steipete (Peter Steinberger) is one of the most-installed utility skills on ClawHub. The concept is straightforward: any URL, file, or video link becomes a distilled summary in seconds. No setup ceremony, no service accounts required, just a command.

The Problem It Solves

When you're working with Claude, the bottleneck is often getting content into the conversation. You have a 40-page PDF, a 90-minute YouTube talk, a competitor's long-form article. You could read it yourself and summarize manually. You could paste a URL and hope the agent can fetch it. Or you can just run summarize and get the essential points in seconds, in whatever format you need.

The friction isn't just the reading — it's that different content types need different handling. PDFs need extraction. YouTube needs transcript access. Some sites block direct fetches. summarize handles all of these paths with a consistent interface.

How It Works

The CLI calls a multimodal model with the content you give it. Depending on the input type:

• URLs: Fetches and extracts the page text (with optional Firecrawl fallback for paywalled/JS-heavy sites)
• PDFs: Extracts and passes text through the model
• Images: Passes directly as vision input
• Audio: Transcribes and summarizes
• YouTube: Uses transcript extraction, with optional Apify fallback for restricted videos

The default model is google/gemini-3-flash-preview — fast, cheap, and good at summarization. You can swap in any provider's model with --model.

Core Commands

Basic Usage

# Summarize a URL
summarize "https://example.com/article"
 
# Summarize a local file
summarize "/path/to/report.pdf"
 
# Summarize a YouTube video
summarize "https://youtu.be/dQw4w9WgXcQ" --youtube auto

Choose Your Model

# Use a specific model (provider/model-name format)
summarize "https://example.com" --model openai/gpt-4o
summarize "https://example.com" --model anthropic/claude-opus-4-6
summarize "https://example.com" --model google/gemini-3-flash-preview

Set once in config:

// ~/.summarize/config.json
{ "model": "openai/gpt-4o" }

Control Summary Length

# Short brief (default)
summarize "url" --length short
 
# Detailed summary
summarize "url" --length long
 
# Full-length (good for dense technical docs)
summarize "url" --length xl
 
# Exact character count
summarize "url" --length 1500

Machine-Readable Output

# Get JSON for piping to other tools or agent processing
summarize "url" --json

Extract Text Only (No Summarization)

# Just extract the content, skip model call
summarize "url" --extract-only

Useful when you want to do something with the raw text yourself.

Handling Tricky Sources

Paywalled or JS-Heavy Sites

Some sites block standard HTTP fetches. Firecrawl is the fallback:

# Auto-fallback to Firecrawl if needed
summarize "url" --firecrawl auto
 
# Always use Firecrawl (for known-difficult sites)
summarize "url" --firecrawl always

Set FIRECRAWL_API_KEY in your environment for this to work.

YouTube with Transcript Restrictions

# Auto-fallback to Apify if YouTube restricts transcript access
summarize "https://youtu.be/..." --youtube auto

Requires APIFY_API_TOKEN. Without it, falls back to available transcript endpoints.

API Key Setup

# Choose one provider
export OPENAI_API_KEY="sk-..."
export ANTHROPIC_API_KEY="sk-ant-..."
export GEMINI_API_KEY="..."
export XAI_API_KEY="..."

No key required if you use the default google/gemini-3-flash-preview model via a configured provider. The skill reads from standard provider environment variables — no summarize-specific auth needed.

How to Install

clawhub install summarize

Or via Homebrew:

brew tap steipete/tap && brew install steipete/tap/summarize

Or run without installing at all:

npx summarize "https://example.com"

Comparison: summarize vs Alternatives

Practical Tips

1. Use --length xl for technical docs — The default length is tuned for articles. For API docs, whitepapers, or long reports, bump to xl or specify a character count to avoid missing key sections.

2. --extract-only as a preprocessing step — If you want Claude to do the analysis rather than the CLI, use --extract-only to dump the text, then pass it in the conversation context.

3. JSON output for pipelines — In any script where you're processing multiple URLs, --json gives structured output you can parse reliably. Combine with jq for filtering.

4. Firecrawl for media sites — News sites, Substack, and many corporate blogs block standard fetchers. Set FIRECRAWL_API_KEY and use --firecrawl auto to handle these transparently.

5. Model cost tradeoff — gemini-3-flash-preview is fast and cheap. For nuanced or technical content where the summary quality matters more, switch to a stronger model like gpt-4o or claude-opus-4-6.

Considerations

• API key required for most models — The default model is free to run if you have a Gemini API key, but you need some key to use the tool.
• YouTube reliability varies — Direct transcript access works for most public videos; --youtube auto with Apify is the reliable fallback for restricted content.
• Firecrawl is a paid service — The fallback for difficult sites costs money. It's optional and only needed for specific paywalled/JS-heavy sources.
• Context window limits apply — Very long documents get chunked. For extremely long PDFs, the summary quality depends on how well chunking preserves the structure.

The Bigger Picture

summarize is one of those tools that, once you have it, changes how you work. The question "what's this about?" becomes a command rather than a task. Research that took 20 minutes of skimming takes 10 seconds. Competitive analysis across 10 company blogs becomes a morning script rather than an afternoon project.

At 30,000+ installs, it's become a default productivity layer for ClawHub users — the kind of tool that quietly multiplies what you can get through in a day.

View the skill on ClawHub: summarize