菜单
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.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Guide users through progressive lessons to learn AI assistant basics
Load context from previous sessions so you can pick up where you left off
Save what you worked on so your AI can pick it up next time
Example skill demonstrating the skill structure. Use as a template for creating your own skills.
Improve LLM prompts using prompt engineering best practices. Invoke when user wants to optimize prompts, improve slash commands, or apply prompt engineering.
Summarize meeting transcripts or notes into actionable format. Invoke when user provides meeting transcripts, notes, or wants actionable meeting summaries.
| 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 valuable insights from content while filtering out marketing hype, self-promotion, and fluff.
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:
What gets filtered:
User provides:
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.
Step 1: Extract Ideas and Recommendations
From the content, extract:
Step 2: Analyze Claims for Hype
Analyze the content for:
Include:
Exclude:
# 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 runs: /fabric-extract-nuggets https://youtu.be/abc123
You respond:
No ingestion by default - just show the nuggets. User can manually save if valuable.
## 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
No actionable content found:
Transcript unavailable:
Ambiguous claims:
Philosophy: Most content is 90% noise, 10% signal. This command finds the 10% and throws away the rest.