fabric-extract-nuggets

name	fabric-extract-nuggets
description	Extract actionable content and filter hype/marketing from videos and articles. Invoke when user provides content with potential hype/marketing, wants signal without noise, or needs bullshit filtering.
allowed-tools	["Bash","Read","WebFetch","Task"]

Extract Nuggets - Content Extraction Without Bullshit

Extract valuable insights from content while filtering out marketing hype, self-promotion, and fluff.

Purpose

You've identified a video or article that might have useful information, but it's wrapped in marketing language, self-promotion, or excessive hype. This command extracts only the actionable nuggets and filters the bullshit.

What gets extracted:

• Actionable steps and techniques
• Specific tools and technical details
• Verifiable facts and data
• Genuine insights (not marketing claims)

What gets filtered:

• Marketing hype ("insane", "game changer", "revolutionary")
• Self-promotion and lead generation
• Repeated concepts and redundancy
• Logical fallacies and unverified claims
• Fluff and filler content

Input

User provides:

• URL (YouTube video, blog post, article)
• OR local file path
• OR paste text directly

Process

1. Extract Content

For YouTube videos:

# Get transcript using yt-dlp
cd /tmp
yt-dlp --skip-download --write-auto-sub --sub-lang en --convert-subs vtt -o "%(id)s" "[URL]"

# Read VTT file and clean formatting
cat [video_id].en.vtt | grep -v "^WEBVTT" | grep -v "^Kind:" | grep -v "^Language:" | grep -v "^[0-9][0-9]:" | grep -v "^$" | grep -v "<c>" > transcript.txt

For web content: Use WebFetch to get article text.

For local files: Read file directly.

2. Extract and Filter

Step 1: Extract Ideas and Recommendations

From the content, extract:

• IDEAS: Key concepts and techniques mentioned
• RECOMMENDATIONS: Actionable steps
• FACTS: Verifiable technical details
• REFERENCES: Tools, technologies, resources mentioned

Step 2: Analyze Claims for Hype

Analyze the content for:

• Unsubstantiated claims (marketing hype)
• Logical fallacies
• Exaggerations and superlatives
• Self-promotional language

3. Filter and Synthesize

Include:

• Ideas with specific details (not vague concepts)
• Recommendations that are actionable (not "sign up for my course")
• Facts that are verifiable
• Tools/technologies with clear use cases
• Technical steps and configurations

Exclude:

• Claims that are unverifiable marketing
• Content with logical fallacies
• Vague statements ("it's powerful", "game changer")
• Self-promotion ("check out my program")
• Repeated ideas (consolidate duplicates)

4. Output Format

# Extracted Nuggets: [Title]

**Source**: [URL]
**Type**: [YouTube/Article/PDF]
**Content Quality**: [Assessment based on claim analysis]


## Actionable Steps

[List concrete, specific steps]
- Step 1: [Specific action]
- Step 2: [Specific action]

## Technical Details

[Specific technical information, configurations, requirements]
- Tool: [name] - [what it does]
- Configuration: [specific settings]
- Requirements: [what you need]

## Key Concepts

[Ideas that passed hype filter - specific, verifiable]
- Concept 1: [Brief explanation with specifics]
- Concept 2: [Brief explanation with specifics]

## Tools & Resources

[Mentioned tools, technologies, references]
- [Tool name]: [Purpose and use case]
- [Resource]: [Why it's relevant]

## Caveats

[Claims that seem questionable or need verification]
- [Claim]: [Why it's questionable]


**Filtering Summary**:
- Original content: [word count]
- Hype/fluff filtered: [percentage]
- Actionable nuggets: [count]

User Experience

User runs: /fabric-extract-nuggets https://youtu.be/abc123

You respond:

1. "Extracting content and filtering bullshit..."
2. [Process content]
3. [Output filtered nuggets in markdown]
4. "Extracted X actionable items, filtered Y% marketing fluff"

No ingestion by default - just show the nuggets. User can manually save if valuable.

Examples

Good Extraction

## Actionable Steps
- Create API key at platform.example.com
- Set environment variable: API_KEY="your-key-here"
- Use `model-name` for best results

## Technical Details
- Model: Example Model 2.0
- Feature: Multimodal (text + voice)
- API endpoint: Uses standard REST API

Filtered Out

• "This is INSANE and a complete game changer!"
• "Join my course to learn more (link in description)"
• "You won't believe what happens next"
• Generic statements with no specifics

Error Handling

No actionable content found:

• Report: "No actionable technical content found. Content appears to be pure marketing/promotion."
• Show brief summary of what was filtered out

Transcript unavailable:

• Report error
• Suggest manual paste of content

Ambiguous claims:

• Include in "Caveats" section
• Note that verification needed

Philosophy: Most content is 90% noise, 10% signal. This command finds the 10% and throws away the rest.