You've Been Using AI the Hard Way (Use This Instead)

Using AI in a terminal beats using it in a browser because it puts project context and files under the user’s control—no more juggling dozens of chat windows, losing context, or rebuilding the same “research” from scratch. The core workflow centers on local context files (like a Gemini “gemini.md” file) that the AI reads every time a new session starts, so the system can pick up where work left off without re-explaining everything.

The transcript walks through a practical setup with Gemini CLI. After installing Google Gemini CLI, the user logs in with a Google account and asks for a task (e.g., coffee research). The terminal experience reveals two advantages the browser hides: a visible context window and the ability to write files directly on the computer. A key demo has the AI produce a markdown file (Best Coffee Method.md) and then generate a project context file (gemini.md) by scanning the working directory. That file becomes the “source of truth” for future sessions. When a new Gemini session starts in the same folder, it automatically loads gemini.md, letting the user request new work (like writing a blog intro) without re-supplying context. The result is fewer fragmented chats and a single, persistent project memory stored locally on disk.

The same pattern expands beyond Gemini. Cloud Code (Claude in the terminal) adds multi-agent capabilities: the user can spawn “agents” that operate with fresh context windows, delegate tasks, and keep the main conversation from bloating. A demo creates a “home lab research expert” agent, then references it from a prompt so it can run web research and compile results. The transcript emphasizes that agents act like separate workers with their own context budgets, which helps avoid runaway token usage and reduces bias from long, one-thread conversations. It also shows “dangerous” modes and headless execution to run multiple tools and agents in parallel.

To keep different AI tools synchronized, the transcript describes a file-based standard: Claude, Gemini, and Codex-style agent context files are kept aligned in the same directory (e.g., Claude MD, Gemini MD, and Agents MD). The user runs multiple terminal AIs simultaneously—each with different strengths (deep work vs. high-level analysis)—while sharing the same local context. The work stays portable because it lives in a folder on the hard drive, not inside a locked browser session.

For project hygiene, an “agent script session closer” automates end-of-day cleanup: it summarizes progress, updates session logs, refreshes context files, and commits changes to a GitHub repository. That creates an audit trail of decisions and makes it easier to resume later.

Finally, the transcript introduces Open Code, an open-source terminal AI tool that can use local models or integrate with a Cloud Pro subscription, including features like session sharing and timeline-style restoration. The overall message is that terminal-first AI turns context into something the user owns—enabling faster iteration, safer organization, and a workflow that can scale from single prompts to multi-agent projects.