Agent Browser Skill (Core)

Purpose

Provide an advanced, production-ready playbook for using agent-browser to automate web tasks via CLI and structured commands.

Quick orientation

• Read references/agent-browser-overview.md for install, architecture, and core concepts.
• Read references/agent-browser-command-map.md for command categories and flags.
• Read references/agent-browser-safety.md for high-risk controls and safe mode rules.
• Read references/agent-browser-workflows.md for recommended AI workflows.
• Read references/agent-browser-troubleshooting.md for common issues and fixes.

Required inputs

• Installed agent-browser CLI and browser runtime.
• Target URLs and workflow steps.
• Session or profile strategy if authentication is required.

Expected output

• A clear command sequence and operational guardrails for automation.

Operational notes

• Snapshot early, act via refs, then snapshot again after DOM changes.
• Use --json for machine parsing and scripting.
• Use waits and load-state checks before actions.
• Close tabs or sessions when done to release resources.

Safe mode defaults

• Do not use eval, --allow-file-access, custom --executable-path, or arbitrary --args without explicit approval.
• Avoid network route, set credentials, and cookie/storage mutations unless the task requires it.
• Allowlist domains and block localhost or private network targets.

Security notes

• Treat tokens and credentials as secrets.
• Avoid --allow-file-access unless explicitly required.

Session & auth playbook

• Session Isolation: By default, agent-browser launches an independent browser instance and does NOT share the user's existing login sessions.
• Reusing Profiles: Use --user-data-dir to persist and reuse login states across runs.
• Connecting to Existing Chrome: Use --cdp to attach to an already running Chrome instance (useful for bypassing complex auth).
• Auth Fallback: For complex OAuth or Cloudflare CAPTCHA scenarios that block automation, immediately hand-off to the user for manual intervention rather than infinitely retrying.

Complex task orchestration

• Step-by-Step Execution: For complex tasks (e.g., batch downloads, multi-page scraping, complex forms), first output a clear plan, then execute step-by-step, reporting the result of each step before proceeding.
• Pacing and Human-likeness: Control operation pacing. Introduce random delays or simulate human input speed to avoid triggering anti-bot mechanisms.
• Sensitive Operations: Explicitly ask for user confirmation before executing sensitive actions such as entering passwords, initiating batch downloads, or deleting data.

Result verification

• Data Validation: If scraped or extracted fields return empty, zero (0.00), or abnormal values, proactively prompt the user to confirm the data source or page state.
• Write Operation Re-check: After executing any write operation (click, fill, submit), you MUST re-snapshot or use get to verify the DOM state has actually changed.
• No Fake Completions: Strictly prohibited to output "Task Completed" or "Updated Successfully" if the underlying data or page state has not substantively changed.