Computer use already existed. Astra made it routine.
Three numbers from the source material. Two are from OpenAI, measured on a benchmark. The third is from the video, measured in two searches, and the kit itself warns: small sample.
ScreenSpot Pro: find the right control on the screen. The previous model scored 76.9%.
OSWorld 2.0: completing real tasks on a computer. The previous version scored 65.7%.
A flight search: pure computer use versus the tool it generated. Two measurements, not a law.
What you’ll know how to do
Write the four-part contract
Exact app, result, scope, and proof of completion. The template all five prompts follow.
Turn a repeated search into a tool
Map a site once without an API using computer use, then come away with a command line that runs with a cheaper model.
Test your MCP or skill as a new user
The agent installs and opens the app, interacts with it, inspects the tool calls, and returns the workflow with evidence.
Have one agent test another
Reproduce your app’s failure through its real interface, connect it to the code, and retest it in the same test.
Take video editing all the way to delivery
Wait for the render, upload it to the editor, apply conservative cleanup, and check the exported file.
Operate your phone from your computer
Apply a setting in the mobile app through the mirroring window and prove it persisted.
Two tracks: the contract, then the five workflows
Each module has six topics, diagrams, at least one copyable prompt with “how to verify,” and a quick three-question quiz.
🧭 Fundamentals and contract
What computer use is, why Astra changed the game, where it fits (browser, desktop, phone), and the four-part prompt contract that gives any workflow proof of completion.
🛠️ The five workflows
Search that becomes a tool, testing MCP inside the app, one agent testing another, video editing through delivery, and the computer controlling the phone. Each comes with a ready-to-use prompt and proof of completion.
People who already use Codex or Claude Code and run into software without an API
Codex / Claude Code user
who wants to use the built-in browser and desktop control with intention, not by trial and error.
People who publish MCPs and skills
and need to simulate someone else’s first use before launch.
Content creator
who renders, edits, and delivers video every week and wants to reduce the manual work.
People who build apps with agents
and want an external agent to audit your app’s own computer use.
More than reading: a learning platform
Everything runs in your browser, with no login. The course remembers where you left off.
Automatic progress
Mark topics as read and track your progress by module, track, and course.
Notes and highlights
Select any passage to highlight or annotate it. It’s saved in your browser.
My journey
Pick up where you left off and see everything you've studied in one dashboard.
Themes and reading
Dark, light, sepia, focus, high contrast. Font size and line spacing.
Quiz by module
Three questions with answers and explanations to check your understanding.
Prompts with proof
Each workflow includes the translated prompt, the placeholders to replace, and what to check at the end.
Where this course’s claims come from
The video
“Computer use by GPT-6 Astra is absurdly good”, with all five cases demonstrated live. The time measurements (77s → 23s, 50s → 25s) come from it.
The field guide and prompt kit
Computer Use Field Guide (Prompt Advisers, 05/09/2026): nine prompts in a text file, the four-part contract, proof of completion, and the google-flights-cli (MIT license). The prompts in this course are adapted translations of those files.
Benchmarks
GPT-6 Astra launch page (OpenAI, 05/09/2026): ScreenSpot Pro and OSWorld 2.0. Documentation for Codex for the built-in browser and computer use tools.
Limitations
Benchmarks don't measure your app. The model name doesn't enable the tool on your account. New google-flights-cli fares are available only inside Codex desktop. Interfaces change, so the prompts ask you to inspect before acting.
Ready to give an agent a mouse and keyboard, with proof at the end?
Start with Track 1. In about 5.2 hours, you'll have the four-part contract and all five workflows ready to adapt to your software without an API.