Use when a task needs a real browser: navigating, filling, clicking, or reading a live page. State the goal and its acceptance criteria for the browser_task tool, then verify the outcome.
Installs into .claude/skills of the current project.
Are you the author of Browser Task?
Add the live security badge to your README. It updates with every re-scan.
[](https://www.skillsdirectory.com/skills/emirbartu-browser-task)
---
name: browser-task
description: Use when a task needs a real browser: navigating, filling, clicking, or reading a live page. State the goal and its acceptance criteria for the browser_task tool, then verify the outcome.
---
# Browser tasks
`browser_task` runs one natural-language goal in a real browser. A Jev policy model, not
you, chooses every click, field value and target, so state the goal and how to tell it
succeeded instead of scripting selectors.
- Reach for it when the work needs a live page: filling a form, walking a site, reading a
rendered page that a plain fetch cannot reach.
- Give it a goal with acceptance criteria, and a start URL when the page matters.
- The returned trace is a report, not proof. Verify the outcome (re-read the page, check the
result) before reporting success.
- Never retry a browser mutation blindly. If the run is `blocked`, report what was observed
and ask how to proceed.
- Keep `max_steps` small for simple goals; each step is one cheap Jev decision.
