browser
the patient hand on the mouse
automation needs an isolated session unless the runtime already provides an authorized selected tab. never create a debugging endpoint for daily Chrome, inspect cookies, or ask for credentials. a sign-in wall is a user action boundary, not a reason to weaken browser isolation.
selectors over coordinatesa screenshot per stepdoes the clicking itself
Use when automating browsers or apps -- agent-browser CLI for testing, screenshots, forms, scraping, cloud browsers (Browserbase), and Electron; operating a user-authorized runtime tab; or driving ChatGPT/Gemini Deep Research. TRIGGER when the user mentions browser automation, screenshots of a page, scraping, a browser account, or deep research.
methodology
- use agent-browser with a named session for ordinary automation, testing, screenshots, and scraping.
- for account-dependent work, use an already-authorized runtime tab first. otherwise follow the global isolated-copy procedure when it is authorized, then verify account state rather than assuming copied state survived. if it is absent, stop for user authentication; never fall back to daily-Chrome debugging.
- route a requested ChatGPT/Gemini research interaction to
driving-deep-research-modeafter the base session is established. - live project
AGENTS.mdand operator directives override these guides.
contents
- trainedlearned layer -- read first
- agent-browserbrowser/app automation with the agent-browser CLI -- testing, screenshots, forms, scraping, Browserbase/cloud browsers, Electron
- driving-deep-research-modedriving a current ChatGPT/Gemini research page through an authorized browser session