AI Knowledge Base (Agent Skill)

English | 繁體中文

🇬🇧 English Documentation

An advanced "Cognitive Accumulation" skill for AI Agent CLIs (Compatible with both Gemini CLI & Claude Code). It is designed to ingest AI tools, URLs, social media posts, and YouTube videos, synthesize them with your existing knowledge, and seamlessly build an automated, token-efficient Obsidian vault. It can also transform this knowledge into beautiful, animation-rich HTML presentations.

🚀 Features

1. Token-Efficient Knowledge Graph: Uses a lightweight _registry.json pattern to search and link your Obsidian notes without passing your entire vault to the LLM. You don't just dump files; you slowly build an interconnected, evolving brain.
2. Interactive Depth Scaling (Shallow / Medium / Deep): The skill actively asks you how deep it should research via interactive menus:
   • Shallow: A quick 100-line summary based only on the provided URL and existing notes. Perfect for "Just tell me what this link is about."
   • Medium: Fetches the URL and performs web searches to verify claims and find counterpoints.
   • Deep: The ultimate analysis. Actively searches for competitors, compares pricing, checks GitHub stars, and builds a comprehensive comparison table.
3. Dual-Mode Video Analysis:
   • Mode A (Knowledge-Centric): Transcribes video audio using locally bundled faster-whisper-srt or cloud-based NotebookLM MCP.
   • Mode B (Visual-Centric): Analyzes video frames via bundled vision_analyzer.py.
4. Subagent Delegation: Intelligently delegates long transcript summarization to subagents (if supported by your CLI) to keep the main context window clean.
5. Frontend Slides Integration: Seamlessly generates standalone, zero-dependency HTML presentations through a "Show, Don't Tell" visual discovery process.

📦 What's Included vs. What You Need to Install

Included in this NPM Package (npm install -g agent-skill-ai-knowledge-base):

• The core instructions.md and SKILL.md logic.
• An automatic install script that configures absolute paths and installs the skill into both your ~/.gemini/skills/ and ~/.claude/skills/ directories.
• Bundled Python scripts: vision_analyzer.py and faster_whisper_srt.py.
• Bundled frontend-slides CSS/HTML templates and export scripts.

⚠️ User Required Installations (Manual Setup): If you do not install the following, the core Markdown extraction and Obsidian linking will still work, but Video processing (Mode A / Mode B) will fail.

1. Python 3.10+ & FFmpeg: Required for audio/video extraction. Must be in your system PATH.
2. Install Python Dependencies: Navigate to the installed skill directory (or the repo root) and run:

   pip install -r requirements.txt

   (This installs yt-dlp, faster-whisper, opencv-python, and python-pptx)
3. NotebookLM MCP (Optional): If you want Google's high-context reasoning instead of Local Whisper, you must configure the NotebookLM MCP server.
   • Step 1: Ensure you have a running MCP server that connects to NotebookLM.
   • Step 2: Add the MCP server configuration to your Agent CLI config file (e.g., ~/.gemini/config.json for Gemini CLI or the respective Claude Code config).
   • Step 3: When the agent asks for your engine choice via the interactive menu, simply select "NotebookLM" and it will route the audio/transcript to your configured MCP tool.

🔄 Workflow Pipeline

graph TD
    A[User Input: URL / Question] --> B{Source Type?}
    B -->|Social Media / JS Heavy| C[r.jina.ai Fetch]
    B -->|Standard Web| D[Direct Fetch]
    B -->|Video / YouTube| E{Video Mode}
    
    E -->|Mode A: Audio| F[yt-dlp extract MP3]
    F --> G{Engine Choice Prompt}
    G -->|Local| H[faster_whisper_srt.py]
    G -->|Cloud| I[NotebookLM MCP]
    H --> J[Noise Reduction / Summarization]
    I --> J
    
    E -->|Mode B: Visual| K[vision_analyzer.py extract frames]
    
    C --> L[Interactive Prompt: Select Depth]
    D --> L
    J --> L
    K --> L
    
    L -->|Shallow| L1[Quick Summary]
    L -->|Medium| L2[Web Search Verification]
    L -->|Deep| L3[Competitor & Pricing Analysis]
    
    L1 --> M[Read Obsidian _registry.json]
    L2 --> M
    L3 --> M
    
    M --> N[Compare Tags/Concepts to Build Links]
    N --> O[Update / Create Note & Registry]
    
    O --> P{Generate Slides?}
    P -->|Yes| Q[Generate 3 Previews]
    Q --> R[User Visually Selects Style]
    R --> S[Export HTML Presentation]
    P -->|No| T[Done]

🛠️ Usage & Trigger Phrases

You can manually trigger the skill in your preferred CLI using /skill ai-knowledge-base, or just paste a link. The agent will automatically detect AI-related intents and present you with interactive choices.

Trigger Examples:

• "Can you analyze this NotebookLM tutorial? [YouTube URL]"
• "Add this to my knowledge base: [Threads URL]"
• "Compare LangChain and CrewAI based on my existing notes."
• "Read this GitHub repo and make a deep dive note: [GitHub URL]"

🙏 Acknowledgments

This skill heavily utilizes the incredible presentation logic from the frontend-slides repository created by Zara Zhang. Note on Open Source Etiquette: The HTML templates, CSS (viewport-base.css), and deployment scripts from frontend-slides have been bundled into this package to ensure a zero-dependency, out-of-the-box experience. All original design philosophies and credit belong to the original author.

🇹🇼 繁體中文說明文件

這是一個專為 AI Agent CLIs (完美相容 Gemini CLI 與 Claude Code) 設計的高階「知識累積」Skill。它可以接收 AI 工具、網址、社群貼文與 YouTube 影片，將其與你現有的知識進行融合，並自動建立一個極度節省 Token 的 Obsidian 知識庫。它還能將這些知識轉換為精美且包含動畫效果的 HTML 簡報。

🚀 核心功能

1. Token 極簡知識圖譜：使用輕量級的 _registry.json 來搜尋與關聯 Obsidian 筆記，無需將整個知識庫塞入 LLM 的 Context。你不是在單純存檔，而是在慢慢建立一個網狀連結的「第二大腦」。
2. 互動式的研究深度選擇 (Shallow / Medium / Deep)： 在處理資料時，Agent 會透過選單主動詢問你希望的研究深度：
   • Shallow (淺層)：僅基於你提供的網址與現有筆記，產出 100 行以內的快速摘要。適合「幫我看一下這個網址在幹嘛」。
   • Medium (中層)：讀取網址後，會額外進行網路搜尋來驗證其說法並尋找反面觀點。
   • Deep (深層)：終極的深度分析。Agent 會主動搜尋競品、比較定價、檢查開源生態系 (如 GitHub Stars)，並為你建立完整的競品比較表。
3. 雙模式影片分析：
   • Mode A (知識深度型)：使用內建的 faster-whisper-srt 或雲端 NotebookLM MCP 將影片轉錄為精準逐字稿。
   • Mode B (視覺動態型)：使用內建的 vision_analyzer.py 擷取影片關鍵影格。
4. Subagent 智能委派：自動將數萬字的長篇字幕丟給子代理人 (若你的 CLI 支援此架構) 進行降噪與總結，確保主 Agent 的記憶體保持乾淨高效。
5. Frontend Slides 完美整合：透過「視覺預覽、不盲選」的互動流程，無縫生成零依賴、支援網頁內建編輯的 HTML 簡報檔。

📦 NPM 套件包含了什麼？你還需要安裝什麼？

本 NPM 套件內建包含 (npm install -g agent-skill-ai-knowledge-base)：

• 核心的 instructions.md 與 SKILL.md 邏輯。
• 一支會在安裝時自動將技能分發到你的 ~/.gemini/skills/ 與 ~/.claude/skills/ 且替換正確絕對路徑的 install.js。
• 打包好的 Python 腳本：vision_analyzer.py 與 faster_whisper_srt.py。
• 打包好的 frontend-slides CSS/HTML 模板與匯出腳本。

⚠️ 用戶需要「手動」安裝的前置作業： 如果你沒有安裝以下套件，純文字網頁抓取與 Obsidian 建檔依然可以運作，但 影片處理 (Mode A / Mode B) 會直接報錯失敗。

1. Python 3.10+ 與 FFmpeg：處理影音必備，且必須加入系統 PATH 變數。
2. 安裝 Python 依賴套件： 請導航至本套件的安裝目錄，並執行以下指令一次安裝所有需要的 Python 模組：

   pip install -r requirements.txt

   (這將會為你安裝 yt-dlp, faster-whisper, opencv-python 以及 python-pptx)
3. 設定 NotebookLM MCP (選用)： 如果你希望使用 Google 強大的高語境推理來取代本地端的 Whisper 轉錄，你必須先設定好 NotebookLM MCP 伺服器：
   • 步驟 1：確保你有一個正在運作且已串接 NotebookLM 的 MCP Server。
   • 步驟 2：將該 MCP Server 的連線設定加入你的 Agent CLI 設定檔中 (例如 Gemini CLI 的 ~/.gemini/config.json 或 Claude Code 對應的設定位置)。
   • 步驟 3：當你傳送影片網址並觸發 Mode A 時，Agent 會彈出互動選單問你要用哪個引擎，此時選擇「NotebookLM」即可。

🔄 系統處理流程 (SOP)

graph TD
    A[用戶輸入: 網址 / 概念] --> B{來源類型?}
    B -->|社群媒體 / JS動態網頁| C[呼叫 r.jina.ai 抓取]
    B -->|標準網頁| D[直接抓取]
    B -->|影音 / YouTube| E{影片解析模式}
    
    E -->|Mode A: 語音知識| F[yt-dlp 下載 MP3]
    F --> G{彈出選單: 轉錄引擎選擇}
    G -->|本地端| H[faster_whisper_srt.py]
    G -->|雲端| I[NotebookLM MCP]
    H --> J[自動降噪與總結]
    I --> J
    
    E -->|Mode B: 視覺動態| K[vision_analyzer.py 擷取影格]
    
    C --> L[互動選單: 選擇研究深度]
    D --> L
    J --> L
    K --> L
    
    L -->|Shallow| L1[快速摘要]
    L -->|Medium| L2[網路搜尋驗證]
    L -->|Deep| L3[競品與定價深度比較]
    
    L1 --> M[讀取 Obsidian _registry.json]
    L2 --> M
    L3 --> M
    
    M --> N[比對 Tags/Concepts 以建立筆記關聯]
    N --> O[寫入新 Markdown 筆記並更新 _registry]
    
    O --> P{是否生成簡報?}
    P -->|Yes| Q[產生 3 款實際 HTML 預覽檔]
    Q --> R[用戶透過視覺挑選風格]
    R --> S[匯出並開啟 HTML 簡報]
    P -->|No| T[完成]

🛠️ 如何觸發與使用 (Usage)

你可以在你的 CLI 中輸入 /skill ai-knowledge-base 來手動啟動，或者直接貼上網址，Agent 發現與 AI 知識學習有關就會自動觸發，並透過終端機選單一步步引導你。

觸發語句範例：

• "幫我把這個 NotebookLM 教學存進知識庫：[YouTube 網址]"
• "這篇 Threads 在講什麼？幫我建檔：[Threads 網址]"
• "根據我現有的筆記，比較 LangChain 跟 CrewAI 的優缺點。"
• "幫我深度分析這個 GitHub Repo 的原始碼架構：[GitHub 網址]"

🙏 鳴謝與開源聲明

本 Skill 在簡報生成階段，深度整合了由 Zara Zhang 所開發的極致美學專案 frontend-slides。 開源運作說明：為了讓用戶能夠透過 npm「開箱即用」而無需手動 clone 其他專案，本套件將 frontend-slides 的 HTML 模板、viewport-base.css 與部署腳本直接打包進了安裝包中。所有關於簡報的美學設計理念與架構歸功於原作者。