A lightweight custom skill lets Claude Code and Codex control your whole screen using only terminal commands, no computer-use harness required.
What is a “drive screen” skill for Claude Code? #
A drive screen skill is a custom instruction file that teaches a coding agent like Claude Code or Codex to control your entire computer, moving windows, clicking, typing, taking screenshots, using only command line tools already on your machine (PowerShell on Windows, AppleScript on Mac, equivalent shell tools on Linux). It works as an alternative to built-in computer-use products, avoiding the need to install a separate harness or out-of-the-box automation tool.
TL;DR #
- A markdown skill file , roughly 400 lines, gives Claude Code or Codex a documented workflow for screen control instead of relying on a dedicated computer-use product.
- No extra software gets installed. The skill drives the OS through native command line tools already present on Mac, Linux, and Windows.
- A companion CLI ships with the skill , packaging up scripts for window discovery, focusing, typing, and pasting so the agent doesn’t have to improvise commands every time.
- Everyday use cases include morning setup routines , opening task managers, browser tabs, and desktop apps automatically, plus testing unfamiliar open-source desktop apps end to end.
- Newer frontier models make this practical. The creator credits models like Claude’s Opus/Sonnet-class releases and GPT-class models with being reliable enough to drive a screen through plain terminal commands, something that wasn’t true before.
- Prompt injection risk exists but is described as lower than expected with current top-tier models, though the creator still recommends sticking to simpler tasks and using the strongest available models.
- The skill is fully customizable , since it’s just a text file and scripts you can edit, unlike closed computer-use tools where tweaking behavior is hard.
Seven tools to build an app. Or just Remy. #
Editor, preview, AI agents, deploy — all in one tab. Nothing to install.
Why build a custom skill instead of using built-in computer use? #
Computer-use tools from major AI platforms, along with various open-source alternatives, already let agents see and control a screen. The problem, according to the creator, is that these tools tend to be heavy, hard to customize, and awkward to fold into an existing workflow. When something breaks or behaves oddly, there’s often no easy way to patch the behavior since the logic lives inside a packaged tool rather than an editable file.
A skill sidesteps that by being nothing more than instructions and helper scripts that a coding agent reads and follows. There’s no separate application to install or maintain. The agent already has terminal access, so giving it a documented way to drive the screen through native OS commands turns out to be enough for a wide range of everyday tasks. That also means the same skill works across Mac, Linux, and Windows, since each OS gets its own set of native commands (PowerShell for Windows, AppleScript for Mac) rather than a single cross-platform binary.
How does the skill actually work? #
The skill is a markdown file combined with a small custom CLI of scripts. The workflow it encodes follows a repeatable loop: discover the relevant window, take a screenshot to understand the current state, focus on the window, then act using one of the documented commands. Scripts handle the fiddly parts, window discovery, focusing, typing, and pasting, so the agent doesn’t have to reinvent the exact syntax for those operations every session. That determinism came from extensive trial and error: the creator describes documenting failure modes observed over many hours of testing and baking the fixes into “hard rules” the agent follows.
The skill also includes logic for watching a session (useful when the agent needs to monitor another coding agent’s session for approvals) and handling “traps,” edge cases specific to the kinds of things being automated, like browser tabs or desktop apps. Because it’s all plain text and scripts, anyone using it can strip out rules that don’t apply to their setup or add new ones for a different Linux distribution or workflow.
Notably, the skill is designed to push back on its own use. Part of the instruction set tells the agent to first consider whether screen control is even necessary, since it’s described as the slowest and least reliable way to accomplish a task compared to something like direct browser automation or an API call.
What can you actually use it for? #
The most immediate use case is a daily setup routine: telling the agent what you’re working on, and letting it open the relevant task management tool, browser tabs, desktop apps, and background services like Docker containers automatically. The creator estimates this saves a meaningful chunk of time every morning.
Remy is new. The platform isn't. #
Remy is the latest expression of years of platform work. Not a hastily wrapped LLM.
A second common use is testing unfamiliar software, particularly open-source desktop apps that don’t expose a browser-based interface for standard automation. In one described case, the agent was given a GitHub repo URL and asked to research it, start the desktop app, and test its features by clicking around and taking screenshots, all without any pre-built computer-use tool. It completed the task end to end using nothing but terminal commands, though the process was noted to be slow and not particularly efficient, characteristics common to computer-use workflows in general.
Is this safe? What about prompt injection? #
Any form of computer use carries some risk of prompt injection, where malicious content visible on screen could trick an agent into taking unintended actions. This is a known concern across computer-use tools generally, not something unique to a custom skill.
The stance taken here is that current top-tier models are meaningfully more resistant to this kind of manipulation than earlier generations, citing research suggesting frontier models are difficult to prompt-inject even without additional guardrails. Based on that, the recommendation is to stick to newer, stronger models when using any computer-use capability, and to keep the tasks handed to it relatively simple, things like environment setup or app testing, rather than open-ended production workflows where a mistake would be costly. For anything closer to production-grade automation, a more heavily guarded, purpose-built harness is still the safer choice.
Is a lightweight skill worth it over a full computer-use tool? #
It depends on what you need. A full computer-use product will generally offer more built-in safety tooling, broader testing, and a more polished interface for complex, high-stakes automation. A lightweight skill trades some of that reliability for simplicity, transparency, and control: you can read every line of what the agent is instructed to do, and you can change it in minutes without waiting on a vendor update.
For routine tasks like setting up a workspace, exploring new software, or automating repetitive clicking-and-checking work, the lightweight approach removes friction: no installation, no separate account, no bloated interface. For anything where a failure could cause real damage (production systems, financial actions, irreversible changes), a more constrained and audited tool is the safer bet.
Frequently Asked Questions #
What operating systems does a drive screen skill work on?
It’s designed to work across Mac, Linux, and Windows by using each system’s native command line tools (PowerShell on Windows, AppleScript on Mac, equivalent shell utilities on Linux) rather than a single cross-platform tool.
Do I need to install anything to use this kind of skill?
No additional software is required beyond the coding agent itself. The skill is a text file plus helper scripts that instruct the agent to use tools already present on your operating system.
Which AI models work best for computer-use skills like this?
The approach depends on frontier-level models capable of reliably interpreting screenshots and issuing correct terminal commands. Newer, stronger models are recommended both for reliability and for reduced susceptibility to prompt injection.
Is prompt injection still a serious risk with computer-use skills?
It remains a known risk in any computer-use setup, but newer top-tier models are reported to be considerably more resistant to it than earlier models. Sticking to simpler, lower-stakes tasks and using the strongest available models reduces exposure further.
How is a custom skill different from built-in computer-use tools?
Built like a system. Not vibe-coded.
Remy manages the project — every layer architected, not stitched together at the last second.
Built-in computer-use tools are typically closed, feature-heavy products that are hard to customize. A custom skill is an editable set of instructions and scripts, making it easier to tweak behavior, fix failure modes, or adapt it to a specific workflow or OS setup.