Ramain Browser Automator

Workflow

When this skill triggers, follow these steps in order.

Step 1 — Capture the task spec

Ask the user for:

• goal — plain-English description of what the browser agent should accomplish
• startUrl — the URL to open first (e.g. https://example.com)
• allowedDomains — one or more domains the agent is permitted to navigate to (prevents scope creep)
• sensitiveDataKeys (optional) — names of environment variables that hold credentials (e.g. LOGIN_EMAIL, LOGIN_PASSWORD); their values will never be hard-coded in the script
• maxSteps (optional) — maximum number of agent steps before the run is aborted (default: 20)

Example spec:

{
  "goal": "Log in to the dashboard and download the latest invoice as a PDF.",
  "startUrl": "https://app.example.com/login",
  "allowedDomains": ["app.example.com"],
  "sensitiveDataKeys": ["EXAMPLE_EMAIL", "EXAMPLE_PASSWORD"],
  "maxSteps": 15
}

If the user is unsure, suggest the worked scenario in examples/input.md.

Step 2 — Save the spec and run the generator

Save the spec as a JSON file, then run from the skill root (skills/ramain-browser-automator/):

node scripts/gen-task.mjs path/to/task.json > path/to/agent_task.py

The engine emits a valid Python script to stdout. Redirect it to a .py file.

The generated script:

• Imports Agent and Browser from browser_use, plus an async LLM client
• Builds a sensitive_data dict by reading each key from os.environ (never hard-coded)
• Configures the browser with allowed_domains from the spec
• Sets the agent task to the goal string
• Caps execution with a max_steps guard
• Wraps everything in an asyncio.run(main()) entrypoint

Step 3 — Tell the user how to run the generated script

The generated script is ready to execute but requires:

1. browser-use installed in a Python environment:

   pip install browser-use
   playwright install chromium
   

2. A supported LLM provider's API key set in the environment (e.g. OPENAI_API_KEY).
3. Any sensitive-data env vars the script reads (listed in the sensitive_data block).

Then run:

python agent_task.py

This skill only generates the script. Network access, a real browser, and an LLM key are required at runtime and are out of scope for this skill.

Step 4 — Iterate

Common refinements:

• Narrow allowedDomains to reduce blast radius
• Add more sensitiveDataKeys to avoid any credential appearing in plain text
• Lower maxSteps for faster fail-fast behavior
• Adjust the goal wording for precision

Update the spec JSON and rerun Step 2 to regenerate.

Example

See examples/input.md for a worked e-commerce checkout scenario and examples/agent_task.py for the generated script.

Run the example yourself:

cd skills/ramain-browser-automator
bash examples/run.sh

The script generates examples/agent_task.py and writes a verification report to examples/output.md.