What an AI agent did and didn't do well rebuilding this tool

The original doctrine tool was a React single-page app I wrote years ago. I rewrote it with Claude, to see how it would go and to test how an AI agent handles legacy software.

What it did well. Claude found the core assets, the translation data and the grid, and rebuilt them against the new constraints in almost no time.

What broke. Quality control for web apps is weak. Even with a test harness, the visual bugs slipped through, and I caught them by eye.

Was it worth it? Not as a product: it already worked and gets under ten visitors a month. As an experiment, yes. It compressed the work to about a seventh of doing it by hand.

Claude needed supervision. The result is one page, no framework, no build step, and the decisions are recorded.