The Terminal Workbench

1. The Philosophy: OS-Level Intelligence

Browser-based AI is limited by "Chat Isolation." Your project files live on your drive, but your intelligence lives in a tab. By moving AI to the Terminal, we achieve:

• Context Persistence: The AI "sees" your entire project folder.
• File Manipulation: The AI can read, write, and create files directly.
• Multi-Agent Sync: Different models (Gemini, Claude) can work on the same shared context files simultaneously.

2. The Tooling Ecosystem

A. Gemini CLI (The Architect)

Best for: High-context planning and free tier usage.

• Command: gemini chat
• Key Feature: Context Anchoring. Run /init to create a gemini.md file. This file acts as the "Project Charter," instructing the AI on the project's goal every time you open a session. No more re-explaining yourself.

B. Claude Code (The Engineer)

Best for: Complex coding tasks and Agent Delegation.

• Command: claude
• Key Feature: Agent Delegation. Use /agents to spin up sub-agents (e.g., "Research Specialist", "Code Reviewer"). Claude can delegate a task to a sub-agent to keep your main context window clean.
• Cost Transparency: Use /cost to see exactly how many tokens your session has consumed.

C. OpenCode / Open WebUI (The Localist)

Best for: Privacy and offline work.

• Configuration: Connects to local models (Llama 3.2, Mistral).
• Usage: Ideal for drafting sensitive research or working without an internet connection.

D. GitHub Copilot CLI (The Agent)

Best for: Integrated GitHub workflows and iterative coding.

• Command: copilot
• Key Feature: Plan Mode. Press Shift+Tab to have Copilot architect a solution before writing code.
• Safety: Requires manual confirmation for shell commands unless configured otherwise.

3. The "Unified Context" Workflow

The Pro-Move: Using multiple intelligences on a single project state.

Create a single context.md (or keep gemini.md and claude.md synced). This allows you to:

1. Have Gemini draft the specific outline in outline.md.
2. Have Claude review outline.md and critique it.
3. Have CodeEx/OpenCode summarize the changes.

All agents read from the same file system, effectively collaborating without direct API connections between them.

4. Advanced: The "Brutal Critic" Agent

The terminal allows for specific "Persona Injection" via flags or config files. We recommend setting up a Review Agent.

/agent create brutal_critic
System Prompt: "You are a ruthless academic reviewer. Your goal is to find logical fallacies, weak citations, and Taylorist assumptions. Never compliment, only critique."

Run this agent against your draft (`draft.md`) to get an unvarnished analysis before human review.