Where does the visual-edit handoff to AI coding agents break?

I’m building VibeQuill, a Chrome extension that lets you select or annotate an element in a website preview and send the requested change, along with page context, to a local AI coding agent through MCP.

I want to understand how developers handle that handoff today. If you’ve used Cursor, Claude Code or Codex for visual UI changes, which part causes the most rework: identifying the right element, describing the change, or checking the result? What do you use now—screenshots, file paths, browser tools, or something else?

Concrete experiences, including cases where this isn’t a problem, would help. If you prefer a short survey, here’s an optional 1-minute one: https://forms.gle/3tirRByg15ErQ9P36 . Please answer only from your own experience.