We're Using AI Wrong: The Chat Window Is a Poor Work Interface

Summary

Ken St. Clair argues that endlessly scrolling chat windows are a poor interface for sustained work because answers, tool activity, and earlier decisions quickly disappear from view. He presents Lahe, an open-source passion project called the Live Agentic HTML Editor, which combines a locally running server with injected JavaScript to place an AI agent inside a local HTML document. Users can collapse the agent rail, select text, request edits through comments and hotkeys, and keep the document itself as the primary workspace. The article describes using this approach to iterate on a website, including more than 700 comments and edits, 97 review pages, 49 logo options, over 85 stock images, 19 line drawings, 17 video files, and more than 120 subagents. The same pattern can extend to todo lists: an item can ask the agent to draft an email or trigger an action without opening a separate chat. One orchestrator can also manage multiple browser tabs, workshop a title in one tab, and write the chosen result back into the main document. Voice transcription enables spoken instructions, while the HTML format supports documents, presentations, and webpage designs. Lahe also records human edits that can be exported as session diffs, creating data for studying and refining an author’s tone. St. Clair frames the project as an open-source tool for reducing task switching and invites others to customize it.