Summary
Cole Medin presents “drive screen,” a lightweight computer use skill he built for Claude Code and Codex that lets an AI agent control an entire desktop without installing a full computer use harness. His argument: models like Claude Fable 5.1 and GPT-6 Astra are now capable enough to drive a screen using only command-line tools, making heavyweight platforms like Anthropic’s built-in computer use or open-source alternatives unnecessary for most tasks.
The video walks through practical daily use cases — morning setup routines that open browser tabs, launch Docker containers, and start desktop apps (saving Medin 10–15 minutes per day), and testing open-source projects with visual UIs that browser automation can’t reach. A live demo shows the skill validating its environment, checking permissions, and completing a setup task in about a minute. Medin also explains that the skill packages documented failure modes and workflow guidance he discovered through iteration, along with helper scripts to make behavior more deterministic — tested across Mac, Linux, Windows, and multi-monitor configurations.
Security tradeoffs between a skill-based approach and a full harness are addressed directly. For developers who want practical computer use integrated into their coding agent without committing to a managed platform, this offers a concrete, self-contained implementation with clear guidance on where the boundaries are.
📺 Source: Cole Medin · Published September 15, 2026
🏷️ Format: Tutorial Demo
