AI Productivity Secrets: My Workflow, Tools & Learning (2024)

My AI Toolkit: A Deep Dive into Productivity

In the past year, I've introduced numerous AI products and methods for enhancing personal productivity. Now, I want to share the specific AI tools I use daily and how I leverage them. This video will unveil my AI workflow, focusing on three key areas: AI-powered search and conversation, my AI knowledge base setup, and how I use AI for learning.

Important Disclaimer

Before we begin, it's crucial to acknowledge the rapid evolution of AI models and applications. My personal knowledge and productivity system is constantly evolving. The tools and techniques I'm sharing are accurate as of April 9, 2025. Any future changes will be promptly updated and shared within our community.

AI-Powered Search and Conversation

Perplexity and API Access

I previously mentioned canceling my Perplexity subscription. While I still support their product, I've transitioned to accessing their models through their API. Perplexity's models, refined and enhanced with search and reasoning capabilities (like their Reasoning Pro model), are incredibly useful.

Raycast: My Go-To AI Tool

To interact with these models via API, I use Raycast. I highly recommend this affordable (around $8/month) software. If I could only keep one AI-related tool, it would be Raycast.

• Accessing AI Chat: I've bound a shortcut (Option+X) to open Raycast's AI Chat. I've set the default model to Sona Reasoning Pro.

• Model Selection: Raycast offers various models with the basic subscription, including Reasoning Pro, Lama 4, Mistro, DeepSeek R1, and Gemini 2.0 Flash.

• Premium Models: For an additional $8/month, you can access top-tier models like Gemini 2.5 Pro and Cloud 3.7.

• Quick Queries: I primarily use Raycast for quick question-and-answer scenarios, making Reasoning Pro an ideal choice due to it's efficiency in answering web search prompts.

Example Usage

Let's say I ask, "Do Perplexity models support web search?" Raycast, using the Reasoning Pro model, will:

1. Perform a web search.
2. Engage in reasoning.

The response will inform me that Perplexity models generally support web search, depending on the model type and version, along with relevant sources. This API-driven approach provides a similar experience to Perplexity's application, but at a lower cost.

Saving Money

By using Raycast with API access to Perplexity's models, I avoid the $20/month subscription fee for the Perplexity application. My other subscriptions include Gemini and Grok (obtained when subscribing to Twitter Blue), and the $8 Raycast subscription is sufficient for my daily AI needs. I also have GitHub Copilot for a yearly price and a $20 subscription to Cursor. With these subscriptions adding up, I have eliminated Perplexity as an optional expense.

Raycast: Beyond AI Chat

I've assigned Option+Z as a shortcut to send selected text as context to AI Chat. This eliminates the need to copy and paste text manually.

Configuring Raycast

Raycast's settings (accessed via Command+,) allow for customization. In the AI section, you can select your default AI model. In the Extensions section, under Raycast AI, you can configure shortcuts for AI Chat (Option+X) and sending selections to AI Chat (Option+Z). Other functionalities include sending screenshots to AI Chat, which is especially useful with models that support image reading, such as Gemini.

Raycast: An Efficiency Hub

Raycast is a powerful efficiency center, and its AI Chat functionality alone justifies the $8/month subscription.

My AI Knowledge Base: Modular and Local

I've recently revamped my knowledge base setup, abandoning previous tools and plugins in favor of a more flexible solution.

Introducing Miiverse and MCP

I now use Miiverse to create a local vector database. This database stores my indexed video scripts (nearly 100 episodes). I then connect Miiverse to AI clients (primarily Cursor) using its MCP (Modular Computation Protocol). This setup effectively implements RAG (Retrieval Augmented Generation).

The Benefits of a Modular Approach

Unlike integrated knowledge base products, my approach separates indexing/retrieval and model generation.

• RAG Breakdown: RAG consists of two parts: Retrieval (indexing and retrieval) and Generation (model-based generation).

• Traditional Integration: Previous knowledge base tools bundled both parts into a single product (e.g., Perplexity Spaces).

• The Problem: Retrieval technology evolves slowly, while models advance rapidly. I want to keep my indexed data separate from the model.

Why Separation Matters

The speed of technological advancements is staggering with Retrieval technologies evolving at a slow pace while models themselves are constantly improving. By separating the two modules, I retain a stable local indexing database.

• Static Retrieval: The indexing and retrieval module remains consistent, stored locally for privacy and security.

• Dynamic Generation: I can freely switch between cloud-based and local open-source models, adapting to the latest advancements.

• Flexibility and Control: This modularity provides greater freedom, control over my data, and enhanced privacy.

Remote Access

My local Miiverse database can be accessed remotely via the cloud, allowing me to leverage it from my MacBook Pro while traveling.

Demonstration with Cursor

Using Cursor as my MCP client (paired with Cloud 3.7 Sonnet), I can query my knowledge base.

• Example Question: "Using Miiverse MCP, how do I deploy a large model locally?"

• MCP Process: Cursor uses MCP to identify available databases and collections within Miiverse.

• Retrieval and Response: It searches the "New Type" collection for relevant keywords and retrieves information from my video scripts, providing detailed instructions for local model deployment.

Agent Trends

The AI is able to use contextual clues to answer my questions more effectively.

• Example Question: "What are the agent trends this year?"

• Efficient Search: Cursor recognizes the context from the previous question and directly searches for relevant keywords.

• Comprehensive Response: It provides a detailed overview of agent trends, including task orchestration, autonomous planning, powerful core models, atomic tools (accessed via MCP), multi-agent systems, and integration into workflows.

Installing Miiverse

Installing Miiverse is straightforward.

1. Visit Miiverse.io and navigate to the documentation.
2. Follow the instructions for Miiverse Standalone installation (using Docker).
3. Run the provided commands in your terminal.

I personally use Obstack, a lighter alternative to Docker, to run Miiverse.

Miiverse Web UI

You can access the Miiverse web UI locally at 127.0.0.1 to view your collections and databases.

Data Import

Importing data into Miiverse can be done through a Python script. I use Cursor to generate this script, which automatically adds all Markdown files from my Obsidian vault to the "New Type" collection.

Cloud-Based Alternatives

While I prefer a local setup, you can also host your knowledge base on cloud platforms like AWS. These platforms offer knowledge base products and MCP services, allowing you to access your data from anywhere using various models.

AI-Powered Learning

AI is revolutionizing how we learn. To keep up with the fast-paced changes, you need a way to learn faster and more effectively.

The Power of AI-Assisted Learning

I embrace AI for learning, starting with Google's NotebookLM.

• Traditional Learning: Linear progression through materials.

• AI-Enhanced Learning: AI extracts the underlying logic and creates a "roadmap" or "network" of knowledge.

The AI Navigation Analogy

Think of a traditional learning method as navigating a path yourself, and AI as providing turn-by-turn navigation to help you.

• AI as Navigator: AI acts as a navigator, guiding you through the knowledge network.

• Flexibility: You can start at any point and explore the material in a non-linear fashion.

NotebookLM: The Best Tool

NotebookLM is currently the best tool for this purpose. I subscribe to Google One to access Gemini and, consequently, NotebookLM Plus.

Demonstration

Let's create a new notebook in NotebookLM. We'll import a PDF, in this case a paper on tariffs.

• Document Analysis: NotebookLM analyzes the document.

• Starting Point: You can start learning from any point that interests you.

Mind Map Generation

The first thing I do is generate a mind map. This mind map provides a visual representation of the document's logical structure. The ability to create a mind map has been added recently, in the last few months.

• Logical Structure: The mind map displays the logical connections between concepts.

• Interactive Exploration: You can zoom in and out to explore the entire structure or focus on specific areas.

• Node-Based Learning: Clicking on a node (e.g., "Correcting Foreign Currency Undervaluation") triggers an AI-generated response based on the document content.

Guided Learning

NotebookLM provides suggested questions based on your current query, guiding you through the logical flow of the information. You can also manually ask questions.

Combining Automation and Control

NotebookLM allows for both automated (suggested questions) and manual (custom questions) learning approaches. By using NotebookLM, I can process and learn from large amounts of complex documents more efficiently.

Overcoming Resistance to AI

Many people are resistant to using AI for learning, fearing that it will diminish their own thinking. However, AI is a tool to enhance efficiency and access more information.

A Competitive Edge

Embracing advanced AI tools like NotebookLM provides a significant advantage in today's rapidly evolving world. Learning how to maximize their utility allows you to keep up with new information faster.

The Power of Consistent Effort

I firmly believe in the principle of "daily incremental progress" – the best way to achieve your goals is to consistently work towards them.

Join the Community

If you're interested in discussing AI, exploring its potential, and connecting with like-minded individuals, join our community.