Our summary
This recipe builds a browser assistant that chooses among listed page actions instead of inventing where to click. Software reads visible links, buttons, and fields, then asks Jev which action best serves the user's goal. Password and file fields are excluded from the offered controls.
The program performs the chosen action using the control's recorded identifier and checks whether the page changed. Separate questions estimate whether the goal is complete or the assistant is stuck. Another text-generating model writes field values when typing is needed; Jev does not generate that text.
In one reported test, both the action choice and completion check accepted the wrong, similar-looking page. An exact destination check exposed the mistake. The author recommends keeping dangerous actions unavailable or behind human confirmation, and treating page content as potentially hostile instructions.
Key takeaways
- Let software define which page actions are available.
- Check the actual destination rather than trusting an AI success signal.
- Require permission before purchases, sending, deletion, or other consequential actions.