Drive-Screen: The Skill That Lets Claude Code Drive Your Whole Computer

No computer-use harness, no special tooling, no repo of infrastructure. Just a coding agent, a shell, and a ~400-line skill. Cole Medin built "drive-screen" to replace the bloated computer-use tools — and it works because the newest LLMs are finally good enough to drive a screen on their own.

Video thumbnail — Claude Code Drives My Whole Computer
🛠️ Cole Medin ⏱️ 13:01 📅 Sep 14, 2026
Computer Use Claude Code Agent Skills Prompt Injection

🖥️ The Drive-Screen Skill 0:00

Cole Medin opens with a confession that doubles as a thesis: he used to be opposed to computer use, but now his coding agent drives his computer every single day — and he doesn't use any tool for it. No Claude's built-in computer use, no Codex computer use, no GitHub harness. He built a single skill called drive-screen, and that's all it took.

The reason the heavy tooling was necessary before, and isn't now, is the models themselves: "large language models, especially with Fable 5.1 and GPT-6 Astra, now they're good enough where they can drive your entire screen for pretty much anything with just the command line." He discovered this by accident a couple of weeks ago and it led directly to the skill.

The pitch: lightweight, easy to customize, cross-platform (Mac, Linux, Windows), and when it makes a mistake the fix is "one more rule in the skill." It sets up his machine in the morning, stages demos, tests desktop apps, and even drives other agent sessions.

The trade-off he's upfront about: he built this because the existing computer-use tools are "really bloated and hard to manage." Powerful, but hard to tweak when little issues come up in your workflow — and that lack of control is exactly what kept him away from computer use as a whole.

🌅 Morning Setup 1:51

The most obvious daily win is the morning setup. You turn on the computer, open your coding agent, and say "get my morning set up — here's what I'm working on today." The agent reads your task-management software over its CLI or MCP, then opens the browser tabs you need, gets Obsidian ready, starts desktop apps and Docker containers — "whatever you need, it can set up everything."

That alone saves "a solid 10–15 minutes every single morning, and that adds up." The key detail is what isn't there: "there is no additional tooling installed here." It's the skill plus simple commands driving everything.

Tested, not claimed: he's run this on Mac, Linux, and Windows, with multiple-monitor setups and different display settings. "It's really good and really adaptable."

🤯 LLMs Can Do Way More Than You Think 2:41

The "big takeaway" he wants to land is broader than the skill itself: LLMs have become far more adaptable than most engineers have updated their mental models for. There are tasks he'd previously written off as "the LLM can't do it" or "you need some massive tool" that now need neither.

His go-to example is video editing: "give it an MP4 and say edit this video, take out the ums, add in B-roll and transitions. Especially with Fable and Astra, it'll knock it out." You probably have a couple of ideas in your head right now for things you assume need a fancy tool — computer use and video editing being the canonical two — and the point is to just try it before buying the tooling.

🔧 How Drive-Screen Started 3:20

The skill was born out of necessity, not a product plan. He was overwhelmed by the volume of open-source projects he was trying — Kestra being one example — many of which have a UI he can't drive with browser automation alone. He wanted computer use to test them, but also wanted something "super lightweight and easy for me to manage."

So he did the obvious experiment: he gave Claude a GitHub URL and said — research the web, dive into the repo to understand the features, then start the desktop app, control my screen to test different features, and leave it open at the end for me to continue testing. He had no idea whether it would work, and half-expected the agent to bail and tell him to do it manually.

It worked end-to-end, "simply by running commands in the terminal." Command after command to click around, screenshot, and analyze results — PowerShell on Windows, AppleScript on macOS, no additional tool. It wasn't fast ("computer use isn't, in general"), but he let it run, worked on another device, and came back to everything done and still open for him to continue.

The lesson from that first session: it didn't start as a skill — just "use PowerShell to drive my screen and do X." But the failure modes started showing up as patterns (the agent getting tripped up, computer use not working through PowerShell in certain cases), so he documented them, laid out the workflow, and built scripts for determinism. That's what became the packaged skill.

⚙️ How the Skill Works 7:41

The whole skill is under 400 lines — "not the longest, but also not the shortest" — and it's structured like a real workflow rather than a prompt fragment. The first thing it does is the least obvious: figure out whether screen control is needed at all. The skill explicitly tells the agent it can push back when a job doesn't require computer use, because there might be a better tool (like browser automation).

The honest framing baked into the skill: "screen control is the slowest and least reliable way to make a computer do something." It's the most adaptable and flexible, but the slowest — and it carries the prompt-injection risk of the agent reading whatever is on screen.

For determinism, it ships custom CLI scripts instead of letting the agent improvise shell commands in real time — packaged best practices for window discovery, focusing, typing, and pasting, built from "hours and hours of testing" into "essentially this little CLI for the agent to call." Below that sit the hard rules (failure modes learned the hard way) and a control loop: discover the window → screenshot for the lay of the land → focus → act, using the CLI. There's even a CLI for watching a session when it's auto-driving another coding agent, checking for approvals and traps.

Customizability is the whole point. On a different Linux distro that needs different commands, you just edit the skill. You can strip traps that don't apply to you or add new ones — "that's the kind of thing you can't really do with a larger computer-use tool."

🔒 Prompt Injection & Security 11:07

Computer use — especially running "just a skill instead of a full harness" — means the agent is reading whatever is on your screen, and that's the classic vector for indirect prompt injection: a malicious instruction embedded in a page or app the agent looks at. Cole links HiddenLayer's writeup 🔗 on exactly this.

Then he makes a claim he admits "might bite me in the butt": you don't have to worry about prompt injection that much anymore, as long as you're on the newest models. His argument is that a growing body of studies shows Fable 5.1 and GPT-6 Astra are "very, very hard to prompt inject" even without fancy guardrails or a harness.

📌 Note (my read, not his): treat that as the author's risk tolerance, not established fact. Prompt injection remains an active research area and a real class of vulnerability — the HiddenLayer article he links is the counterweight. His own hedge is the honest one: use at your own risk, use the most powerful models, and for production-grade computer use he still reaches for a proper harness.

The rest of his security posture is sound and specific: stick to simple things (morning setup, testing desktop apps), keep the model current, and remember the skill is "pretty token efficient — it's not the fastest, but that's because it's waiting on inputs, not burning tens of thousands of tokens."

📦 Install & Verdict 12:22

Installation is deliberately frictionless. You can install it as a Claude Code plugin to pull in all of Cole's skills, or just copy the drive-screen folder into any coding agent — "drop it into any project, use it with whichever coding agent you're on, or install nothing and take the ideas." It's MIT and free.

The skill folder ships with exactly what you'd expect of a well-built agent skill: SKILL.md, a references/ directory, and the scripts/ that make up the custom CLI. If that structure looks familiar, it's because it's the same pattern Hermes and Claude Code skills share — a declarative instruction file plus supporting scripts.

My verdict (checked against the live repo):Worth trying for light automation. The repo (coleam00/skills 🔗) is MIT-licensed, actively pushed (Sep 15, 2026), with 517 stars and 149 forks. The skill is honest about its own ceiling — the author repeats twice that screen control is the slowest and least reliable path — and the security guidance is the right shape: newest models, simple tasks, real harness for production. It earns its place as a zero-dependency on-ramp to computer use, not a replacement for a hardened harness.

💡 Key Takeaways

  1. You don't need a computer-use harness anymore. With Fable 5.1 and GPT-6 Astra, a coding agent plus shell commands (PowerShell / AppleScript) can drive an entire screen reliably.
  2. The skill is the tool. ~400 lines of SKILL.md plus a custom CLI for window discovery, focusing, typing, and pasting — no install, MIT, cross-platform.
  3. First rule: question whether screen control is needed. The skill is built to push back — it's "the slowest and least reliable way to make a computer do something."
  4. Determinism comes from scripts and hard rules, not clever prompts. Failure modes get documented as patterns and encoded into a control loop: discover → screenshot → focus → act.
  5. Prompt injection is the real risk of a harness-free setup. The agent reads whatever is on screen; the author leans on the newest models and simple tasks as his guardrail.
  6. Customizability beats big tools. Fixing a mistake is "one more rule in the skill" — you can't do that with a packaged computer-use platform.
  7. The bigger lesson: update your priors on what LLMs can do. Tasks you've written off (computer use, video editing) may just work now.

🔗 Resources & Links

Source video: youtube.com/watch?v=SWEThyRHMgQ

⏱️ Timestamp Index

0:00 My new favorite skill: Drive Screen
1:51 Morning setup with your agent
2:41 LLMs can do way more than you think
3:20 How drive-screen started
7:41 How the skill works
8:52 The custom CLI scripts
9:57 Hard rules, control loop, traps
11:07 Prompt injection & security
12:22 Install the skill
☰ View all