Skip to main content

Skill System

Understand the two-layer entry Skill and get-skills runtime model.

Written by Maggie

Browser-act uses a two-layer Skill model:

  1. The entry Skill makes the agent aware that Browser-act exists.

  2. The get-skills runtime returns current environment state, command guidance, and dynamic instructions.

Install the entry Skill

Ask your agent:

Install the browser-act Skill from:
https://github.com/browser-act/skills/tree/main/browser-act

After installation, the agent can recognize Browser-act tasks and call get-skills before using the CLI.

Two layers

flowchart TD
  User["Browser task"] --> Entry["Entry Skill"]
  Entry --> Runtime["get-skills runtime"]
  Runtime --> State["Runtime state"]
  Runtime --> Rules["Safety rules"]
  Runtime --> Version["Version guidance"]

Layer 1: entry Skill

The installed Skill file is intentionally small. It:

  • triggers agent awareness

  • contains activation language

  • tells the agent to load runtime content through get-skills

  • changes rarely

Layer 2: get-skills

The first command for a Browser-act task should be:

browser-act get-skills core --skill-version <version>

The output gives the agent:

  • CLI and API key status

  • available browsers and descriptions

  • active sessions

  • core command guidance

  • dynamic instructions for the current environment

Topics

core

get-skills core --skill-version <v> returns the core workflow, commands, browser state, and safety rules for most tasks.

advanced

get-skills advanced returns proxy, profile import, privacy mode, and advanced browser setup guidance.

main

get-skills main returns the latest Skill content when a version mismatch is detected.

Progressive loading

Tip: Most tasks only need get-skills core. Load advanced topics only when the task requires them. This keeps the agent context focused and reduces irrelevant instructions.

Version compatibility

The --skill-version parameter lets the CLI detect mismatches:

browser-act get-skills core --skill-version 2.0.2

When versions are incompatible, the output includes direct upgrade guidance:

CLI too old

Run uv tool upgrade browser-act-cli.

Skill too old

Run browser-act get-skills main.

Dynamic instructions

get-skills adjusts guidance based on the current state:

Situation

Instruction returned

Multiple browsers exist

Choose by desc match or ask the user

Active sessions exist

Respect session ownership and naming

API key is missing

Avoid stealth-only features or guide authentication

Sensitive browser is selected

Ask before opening a confirm_before_use browser

Agent workflow

flowchart TD
  Ask["User asks"] --> Activate["Skill activates"]
  Activate --> Load["Run get-skills"]
  Load --> Read["Read state"]
  Read --> Select["Select browser"]
  Select --> Confirm["Confirm if needed"]
  Confirm --> Run["Run commands"]
  Run --> Finish["Close session"]

Agent compatibility

Browser-act can work with agents that can:

  1. load Skill files

  2. run shell commands

  3. read text output

Known compatible environments include Claude Code, GitHub Copilot, Cursor, Windsurf, Gemini CLI, OpenCode, Codex, and similar agent runtimes.

Learn more

Quick Start

Run the first Browser-act automation loop.

Command Reference

Open the full command index.

Skill Forge

Generate reusable task-specific skills.

Did this answer your question?