Lavish：把 AI HTML 评审变成可交互反馈

Claude writes an HTML artifact. Instead of eyeballing it in a chat window, I open it in an actual local review surface, click the exact element that’s wrong, and queue that feedback for the agent to pick up. It’s called Lavish. A local CLI that turns any agent-generated HTML file into an interactive review page, so you’re not scrolling through raw markdown to catch a mistake. → Opens agent-generated HTML in your own browser, click any element or highlight any text, and it captures the precise target, not just “somewhere in this file” → Edit rendered Mermaid diagrams like a whiteboard, then queue that exact change as feedback for the agent → Runs entirely local. No cloud round-trip to view or annotate an artifact, session state stays on your machine under .lavish-axi/ → Ships built-in playbooks for diagrams, tables, comparisons, plans, code diffs, input collection, and slides, so the agent knows how to structure what it’s showing you before it writes the HTML → Session hooks plug directly into Claude Code, Codex, OpenCode, and GitHub Copilot CLI, so it triggers automatically instead of being a separate tool you have to remember to open → Runs a layout audit before it even shows you the artifact, catching overflow, clipped text, and broken elements before you waste time reviewing a broken render No install required, it runs on demand through npx -y lavish-axi. Every AI coding tool generates plans, diagrams, and comparisons buried in chat text you have to parse manually. Lavish turns that output into something you can actually click on and correct. MIT License. Built by Kun Chen.