I am doing similar experiments for in-browser user testing, where the user is basically Claude. Incredible results with this simple pipeline: 1. Claude tests a feature 2. Notes down all friction and pain points 3. Convert those as a prioritized todo list 4. Use Claude Code or similar to action the todo list

It has some friction in-between steps 3 and 4, but nothing that can't be solved without running `claude --chrome` via CLI instead of the Chrome extension.

It would be neat if it had a headless mode.