AGENTS.md:When doing UI work examine screenshots in both desktop and mobile views before handing over a task and fix any issues.This might do the trick without any complex setup.replied to a while ago:
Thanks! That’s a good idea, but I think it can be done simpler. Codex does screenshots itself when asked to, so I added this to
replied to:I'm not saying you should, but when I'm designing things and want the agent to be able to use its vision facilities - I get it to write Playwright tests and sometimes even chunk the screenshots to isolate specific areas. Last job I was building a kiosk for a client and got my agent to make standalone scripts to get desired output so here's an example you could feed your agent for inspiration if you want: