Youve Been Using AI the Hard W

This video introduces the concept of using AI tools within the terminal environment instead of browser-based applications, highlighting increased speed, control, and advanced features.

The Problem with Browser-Based AI #

• Browser apps are slow and lead to context loss.
• Users often create multiple chats across different LLMs (Gemini, Claude, ChatGPT) and struggle to track information.
• Copy-pasting into note apps is inefficient and leads to disorganized projects.
• Projects become scattered across numerous chats and notes.

Introduction to Terminal-Based AI #

• Terminal versions of AI tools are faster and more powerful.
• AI companies are not widely promoting terminal versions to general users, focusing on developers.
• Terminal AI can be used for various tasks beyond coding, such as writing and research.
• Terminal AI offers better context management and file access.

Gemini CLI #

• Installation:
  • Command: installing Google Gemini CLI (specific command provided in the video transcript)
  • Works on Mac, Windows (via WSL), and Linux.
  • Mac users can use Brew Brew.Install Gemini Dash CLI.
• Basic Usage:
  • Create a dedicated directory (M-K-D-I-R coffee project, CD Coffee Project).
  • Launch with Gemini.
  • Log in with a Google account (free Gmail accounts work).
  • Ask questions, e.g., "How do I make the best cup of coffee in the world?"
• Superpowers:
  • Access to the latest models (e.g., Gemini 2.5 Pro).
  • Displays context window percentage (e.g., "99% context left"), which browser versions hide.
  • File Reading/Writing: Can create and modify files on the computer (e.g., .md files).
  • Access to local files: Can interact with Obsidian vaults or other local notes.
  • Script Execution: Can run Bash and Python scripts.
• Tools:
  • Type /tools to see available actions.
  • Can add more tools.
• Context Management (gemini.md):
  • Type /init to create a gemini.md file.
  • This file analyzes the project, reads the folder and files, and creates context instructions for Gemini.
  • Gemini loads this file for context in every new session within that directory.
  • Users can ask Gemini to update the gemini.md file with thoughts, research, and project progress.
  • This eliminates the need for re-explaining context and keeps projects organized in a single file.
• Ownership: All project data (context, files, decisions) are stored locally on the hard drive, giving users control and avoiding vendor lock-in.

TwinGate (Sponsor Segment) #

• Problem: Traditional VPNs give remote employees broad access to internal networks, posing security risks if AI tools also access the computer.
• Solution: TwinGate offers Zero Trust Network Access.
  • Users are granted access only to explicitly allowed resources (e.g., marketing team accessing NAS but not secret file server).
  • Trusted Profiles: Ensures only secure devices connect.
    • Checks for OS, hard drive encryption, screen lock, firewall, antivirus, OS updates (e.g., blocking Windows 10).
    • Integrates with EDR/MDM solutions like CrowdStrike, Intune, Jamf.
  • Benefits: Enterprise-grade security, free for up to five users, easy setup (five minutes).

Claude Code #

• Primary Tool: Used as the daily driver due to its "agents" feature.
• Cost: Not free, but users with a Claude Pro subscription (around $20/month) can log in and use it without API keys. Claude Pro is recommended if paying for only one AI subscription.
• Installation:
  • Command: NPM install (specific command provided in the video transcript)
• Basic Usage:
  • Launch with Claude in the directory.
  • Log in and grant folder access.
  • Asks questions, e.g., "I need to find a NAS for my house. Here's my budget..."
  • Features a TUI (terminal user interface) and thinking mode (press Tab).
  • Security-first: Asks for permissions for most actions.
• Context Management (claude.md):
  • Type /init to create a claude.md file, similar to Gemini, for context.
  • /context command shows token usage and remaining context.
• Agents:
  • Type /agents to manage agents.
  • Creation:
    • Create agents (e.g., "home lab research expert") for specific tasks.
    • Can be tied to a specific project or be personal (accessible across projects).
    • Can be given full access to tools or restricted.
  • Delegation: Claude can delegate tasks to sub-agents.
    • Each sub-agent gets a fresh context window (e.g., 200,000 tokens), protecting the main conversation from bloating.
    • Allows for parallel task execution without disrupting the main AI's context.
  • Use Cases:
    • Protecting context and avoiding bias.
    • Comprehensive review (e.g., "brutal critic" agent for script review).
    • Running multiple agents simultaneously (e.g., researching network equipment, pizza places, graphics cards).
  • Dangerous Mode: Use dash dash dangerously dash skip permissions when launching Claude (e.g., Claude dash R dash dash dangerously dash skip permissions to resume a conversation) to bypass continuous permission requests.
• Multi-AI Collaboration:
  • Claude can use Gemini for research by running Gemini in headless mode (Gemini dash P followed by prompt).
  • Agents using other AIs: Create an agent that uses Gemini in headless mode for research.
• Output Styles:
  • Type /output style to view or create new output styles.
  • Output styles control the system prompt and persona of Claude (e.g., "script writing" output style).
  • Can be project-specific or global.
• Other Features:
  • Shift + Tab for planning mode (AI creates a plan, user approves).
  • Prompts, hooks, custom status lines, image pasting.

Multi-AI Workflow (Claude, Gemini, Codex) #

• Setup: Open all three AI tools (Claude, Gemini, Codex) in the same directory.
• Context Sync: Ensure all context files (claude.md, gemini.md, agents.md for Codex) are synchronized.
• Parallel Tasks:
  • Example: Claude writes a hook from an authority angle.
  • Gemini writes a hook from a discovery angle.
  • Codex reviews both.
• Benefits:
  • Three different AIs working simultaneously on the same project.
  • Share the same context and can see each other's work (they operate on the same files).
  • No copy-pasting needed.
  • Claude and Gemini are good for deep work; ChatGPT (Codex) is good for high-level analysis.
• Portability: The entire project (files, decisions, context) is contained within a folder on the hard drive, which can be copied anywhere.
• Vendor Lock-in Avoidance: Users own their context, making it easy to switch to new AI tools.

Project Management and Synchronization #

• Closing Sessions: Use a "session closer" agent (personal agent).
  • Summarizes discussions.
  • Updates a session summary file.
  • Updates core project files.
  • Synchronizes claude.md, gemini.md, and agents.md.
  • Commits project changes to a GitHub repo (treating projects like code for version control and history).
• Benefits:
  • Documentation and tracking are automated.
  • Easy to pick up where left off, even if tired.

Critic Agents #

• Purpose: Create agents specifically designed to critique work rigorously.
• Example: "Brutal Critic" agent.
  • Designed to be mean and hard to please.
  • Helps overcome AI's natural agreeableness.
  • Ensures quality by providing unbiased feedback from a fresh perspective.
  • Can leverage external frameworks (e.g., for script writing) for critique.

Open Code (Open Source Alternative) #

• Features:
  • Open-source.
  • Supports any model.
  • Offers free Grok AI access.
  • Allows logging in with a Claude Pro subscription (bypassing API key payment per use).
• Installation: One command installation (specific command provided in the video transcript).
• Usage:
  • Launch with open code.
  • Defaults to Grok code.
  • Local Models: Can edit config/OpenCode.jsonc to use local models like Llama 3.2.
  • Claude Pro Integration: open code auth login allows logging in with Claude Pro subscription.
  • Model Switching: Use /model to switch between models.
  • Sessions: View previous sessions (/sessions), share sessions (/share), view and restore timeline (/timeline).
  • Agents: Supports agents, headless servers, and session export (JSON).

Conclusion and Call to Action #

• Terminal AI provides greater control over context and superior features compared to browser-based AI.
• Encourages users to try Gemini CLI (free) and Open Code (free/local models/Claude Pro).
• Claude Code is powerful but paid.
• Users can add tools and integrate with MCP servers (referenced another video).
• Call to action for YouTube engagement (like, subscribe, comment).
• Questions to the audience: What's your favorite terminal AI tool? Do you care about this?

End Segment Prayer #

• Expresses gratitude for the viewer.
• Prays for joy, removal of anxiety, peace, blessings for families, path forward through difficulties, career excitement, opportunities, and ultimately, a deeper relationship with Jesus.