The Agent Skills Directory

[COMMAND_EXECUTION]: The skill generates JavaScript modules (variant.js) at runtime based on recorded web sessions and executes them using a Node.js runner. It also utilizes shell scripts to orchestrate Playwright and ffmpeg for recording and replaying UI flows.
[DATA_EXFILTRATION]: The skill captures full network activity, including session cookies and authorization tokens, into session.har files. While the process_har.py script redacts these in the lesson summaries, the raw sensitive data remains stored locally and is accessed by the agent for replaying actions, creating a risk of accidental data exposure.
[EXTERNAL_DOWNLOADS]: The scripts perform automatic installation of the playwright package and may download browser binaries and media tools through Playwright's installation commands.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection as it processes untrusted website data to inform its logic and code generation.
Ingestion points: External website responses and network logs captured in session.har and UI interaction steps in actions.js.
Boundary markers: No explicit delimiters or instructions are used to prevent the agent from obeying instructions embedded in the captured web data during script generation.
Capability inventory: The skill possesses capabilities for shell command execution, filesystem read/write operations within the lesson directory, and network communication via Playwright and curl.
Sanitization: A distillation script redacts common secret patterns (e.g., tokens, passwords) using regular expressions, but does not validate or sanitize the overall content of the network logs.

teach-web-actions