Skip to content
Agent Orchestration
Skill

/browser-control

Open, inspect, interact with, and verify websites or local web apps in Poracode's isolated browser. Use for visible page state, navigation, screenshots, console or network evidence, and end-to-end UI testing; do not use it for semantic service operations when a purpose-built

BOOST
From plugin
poracode
11722 skills
Install
$ npx -y skills add Porabuild/Poracode --skill browser-control --agent claude-code

How it fires

How this skill gets triggered: by you, by Claude, or both.

  • Fires itselfAuto-invocation. Claude auto-loads it when your prompt matches the work.Auto-invocation is when the right skill fires by itself at the right moment, driven by a FLOW.md router and a hook, instead of you invoking it by name. It is the difference between a skill being installed and a skill actually getting used.Read the full definition →
  • You can call itInvoke it directly when you want it.
  • Slash command/browser-control

Context preview

The summary Claude sees to decide when to auto-load this skill.

Open, inspect, interact with, and verify websites or local web apps in Poracode's isolated browser. Use for visible page state, navigation, screenshots, console or network evidence, and end-to-end UI testing; do not use it for semantic service operations when a purpose-built

SKILL.md

browser-control.SKILL.md
name: browser-control
description: Open, inspect, interact with, and verify websites or local web apps in Poracode's isolated browser. Use for visible page state, navigation, screenshots, console or network evidence, and end-to-end UI testing; do not use it for semantic service operations when a purpose-built connector is available.

Browser Control

Use Poracode's `browser` MCP when the task depends on a rendered page, visible interaction, or local web app. If the request is really about structured data or a service operation and a purpose-built connector is available, use that connector instead. An explicit request for Poracode's browser wins.

Workflow

1. Call `browser.api` when you need the current API map, then call `browser.enable` once before the first browser action. 2. Reuse a relevant tab from `browser.list_tabs`; otherwise open the exact URL the user supplied or the known local target. Do not guess a remote site or substitute web search when authentication blocks the requested page. 3. Establish the baseline with the current URL plus `browser.snapshot` or `browser.find`. Prefer accessible roles, names, and returned element refs over brittle selectors or coordinates. 4. Use `perform` for a bounded sequence of known actions: `{steps:[{action:"fill",ref:"@e1",text:"Ada"},{action:"click",ref:"@e2"},{action:"wait",text:"Saved"}]}`. It stops on failure and returns one final compact snapshot (`observe:"none"` omits it). Use `fill` for replacement and `type` for appending. Split batches at decisions, new targets, and navigation; do not replay completed steps after failure. 5. Include a condition `wait` at the end of a batch for asynchronous changes, then inspect its returned observation. After navigation, wait for the expected URL, text, or element and inspect the resulting state. For web-app verification, also check relevant console errors and failed network requests. 6. Capture a screenshot when visual layout or appearance is part of the requirement. 7. Call `browser.disable` before asking the user for input, waiting on an external event, or finishing.

Boundaries

  • The in-app browser is isolated from the user's personal Chrome profile. Do not assume it contains existing logins, cookies, or extensions.
  • Never inspect cookies or storage unless the task requires it and the user authorized that data access.
  • A successful click is not proof of success. Verify the user-visible or application state it was meant to produce.
  • Pause before purchases, submissions, messages, deletions, or other irreversible external actions unless the user explicitly authorized that exact action.

Output

Report the tested URL and flow, the final observed state, and the evidence used. Separate visual, console, and network findings, and state any step that could not be verified.

Read more
Ships withporacode

One window for all your AI coding agents. Run Claude, Codex, OpenCode, Gemini, Antigravity, Cursor, and Copilot side-by-side. Terminal and chat, any layout.

Get the whole plugin

Other skills on poracode.