The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The documentation instructs the user to install well-known and reputable third-party packages, including Playwright and official SDKs for Anthropic and OpenAI. It also includes commands to download browser binaries via the Playwright CLI.
[COMMAND_EXECUTION]: The skill uses subprocess.run to execute a bundled Python script (solve_captcha.py). This script uses only the Python standard library to communicate with CAPTCHA solving APIs and is invoked with local environment variables for configuration.
[DATA_EXFILTRATION]: The skill transmits site-specific CAPTCHA metadata, such as the sitekey and pageurl, to external solving services (2captcha, CapMonster, and Anti-Captcha). This is the primary and intended function of the skill to obtain solution tokens.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection as it processes visual data from external websites.
Ingestion points: Browser screenshots of CAPTCHA challenges are captured and sent to external vision AI models (Claude or GPT-4V) via solve_image_grid.py.
Boundary markers: No explicit delimiters are used in the vision prompt to distinguish the untrusted image content from the task instructions.
Capability inventory: The script has the capability to perform automated clicks on browser elements and navigate pages using Playwright.
Sanitization: The skill implements sanitization by using regular expressions to extract a JSON array from the AI's response and strictly validating that the content consists only of integers corresponding to valid grid cell numbers.

captcha