I let AI rebuild a dead Java desktop app for the web. Here's where it broke.
This title could be clearer and more informative.Try out Clickbait Shieldfor free (5 uses left this month).
A developer attempts to rebuild RSSOwl, a classic Java/SWT desktop RSS reader, as a multi-user web app using Vaadin 25 and Claude Code as the AI assistant. The experiment reveals what AI handles well (table sorting, reactive Signals-based selection, context menus, dramatic code reduction) and where it fails (hallucinating APIs from pre-25 training data, silent styling bugs, subtle sorting edge cases, fabricated default values). Key lessons: always point the AI at live framework docs via an MCP server rather than relying on its stale training data, verify behavior by running the app not reading diffs, and expect the AI to nail structure while missing fidelity details that require human domain knowledge. Some desktop features (Eclipse plugin menus, embedded browser, Google Reader sync) have no web equivalent at all. The resulting app — with OIDC login via Keycloak, full-text search, and multi-user feed lists — ended up more complete than planned, but required constant human oversight to catch correctness gaps.
