
The author ran a small experiment with his two AIs. Claude had built two projects for him, both public on his GitHub (ouijabored66): a self-hosted voice assistant called jarvis (an iPhone app plus a server on his own computer) and a Cider 4 internet radio plugin (TuneIn plus Radio Browser). He asked Jarvis — his own personal assistant, Meta's Muse, which he named Jarvis — to go through both repositories line by line and audit Claude's work: leaked secrets, personal data, sloppy code. The verdict was clean: no committed API keys, tokens, or personal data; tokens are generated at runtime; sensitive config files are properly gitignored; the server only listens on localhost; and the Swift code had zero force unwraps, better than he expected. They then filled the gaps: an MIT LICENSE under his own username, author and repository fields completed in plugin.yml, and the full Vue/TypeScript source added to the Cider plugin repo (which previously only had a compiled bundle). One hiccup: his GitHub app had been logged in but not installed, so the first push was rejected; once installed, the push went through.
Summary: paraphrased from the original post; full original text not captured.
Jarvis is the name the author gave his own Muse assistant; one of the projects Claude built happens to be named jarvis too. The auditor is Muse; the audited code is Claude's.
Two AIs, one checking the other: Muse audited Claude's code line by line — and the code was clean.