GitHubWhat they did
Arena used $250 of free credit to build Sudo Play, a small macOS menu bar app in Laravel and NativePHP, entirely through Claude Code cloud sessions started from the terminal. He ran two Claudes. The one on his laptop wrote a brief for each milestone (what to read, branch name, scope, tests, what not to touch) and later reviewed the result. The cloud one, in a Linux container, read the brief, wrote tests first, opened a pull request and stopped. He started each session, pasted the cloud session's final report back into the local one by hand, and merged nothing himself without checking. Lasting rules lived in AGENTS.md: never merge your own PR, stop and ask when the plan and requirements disagree.
How it went
The AGENTS.md rules held in every session. One session tested the planned rotation algorithm over 400 days, found it repeated the previous day's pick 42% of the time, and stopped to ask instead of building to pass the tests. Brief-only rules fared worse: three sessions in a row tried to install PHP 8.4 after being told it could not be done. A CLI change silently dropped the brief, and one session built a whole PR without it. Twice tests were green but the app was wrong, once because every test faked Spotify. Each Laravel session re-solved the same four container problems.
Worth knowing
After the CLI change, the session has to be created first and the brief sent with a second command; Arena now makes the brief's first line ask for the branch name so a lost brief shows up at once. A Generated by Claude footer kept appearing on PR descriptions despite AGENTS.md and had to be removed from his own machine. He still prefers the local tool on a Max plan for cost.
Discord