What they did
Starting from an empty folder on November 15, Phutson used Claude Code as a shared workspace for several personal projects at once: a React component library, an icon set, a Thai font, a portfolio rebuild, a typing trainer, an uptime monitor on Cloudflare Workers and a pixel-art music page. A workspace CLAUDE.md and one per project set the rules, including an ask-before list for pushing, creating repos, publishing to npm and bumping versions. The agent kept its own memory notes on decisions and machine quirks. It set up Changesets, GitHub Actions release pipelines with npm trusted publishing, and branch rulesets. For the icons it wrote a design spec and a linter, drew the first 140 itself, then split the other 299 into batches for smaller sub-agents that could draw but not build or commit, and reviewed their output as contact sheets.
How it went
Git counted about 140 commits in four days: 60 on the component library (v0.8.0 on npm, 25 components), 16 on the icon set (439 icons, v0.1.0 on npm), 13 on the font and the rest spread over the other repos. On the last day one change flowed through four repos, from the icon release to the library to three apps that depend on it. Of the 299 sub-agent icons, 15 to 20 needed fixes even though all passed the linter, such as an umbrella that read as a mushroom at 16 px. The agent also traced a failed npm publish to a placeholder token and a production-only CSS bug to the minifier. Taste, logins, first publishes and every push stayed with him.
Worth knowing
The linter could not tell whether an icon read correctly, so a single reviewer looking at rendered output was what made the parallel drafting usable. Credentials stayed manual by design. Some builds would not run locally because Windows App Control blocks unsigned Python modules, so the agent verified those through CI instead.