Deliver a software feature autonomously (partial)
Claude computer use produced 70 files for a planned feature but the code didn't work, so the author went back to manual work.
Done withClaude computer use
- What they did
- The user discussed the feature deliverables and agreed on a concrete plan with Claude, then granted it full autonomous control to implement the entire feature. Claude generated 70 files on its own without further input or tests. The user reviewed the output afterward.
- How it went
- The code did not work. The user deleted the branch, cleared the session, and returned to using Claude only for incremental pair-programming steps.
- Worth knowing
- The full autonomous run produced unusable output after the agreed plan, prompting immediate abandonment.
Try it yourself with Claude computer use
In my repo at [repo path], implement this feature: [feature description]. First write a short plan and the list of files you'll change, and wait for my okay. Then build it in small steps, running the tests after each step. Stop when the tests pass and I can run [command] to see it working, and don't delete files without asking.
Source: X · Oct 1, 2026
Five of these in your inbox every morning
The best things people got an AI agent to do, each with the prompt to try it.
More like this
Fill out a Form 1040 with computer use
A tester gave Claude Computer Use data stored on disk and checked whether it could find the fields and fill a simplified Form 1040 accurately.
Run content research and sponsor outreach with Claude computer use
The author tested Claude computer use in Cowork and Claude Code over two days on six workflows including research, sponsor outreach and product listings.
Click through macOS app builds to catch visual bugs
A developer used Claude Code computer use to compile, launch, and click through macOS app screens to find layout bugs.
- Claude
Build a C compiler with a team of parallel agents
Anthropic's engineering write-up on tasking Opus 4.6 agent teams with building a C compiler.
Write and debug code across eight test tasks
A reviewer had Manus build a currency converter, fix a find_max bug and make a landing page, with mostly strong results.
Click through apps to verify its own work
With computer use, Devin built and played a desktop game, ordered on Amazon, and tested its own work by clicking through the app.