The Agent Skills Directory

[PROMPT_INJECTION]: The skill uses mandatory and forceful language in SKILL.md to dictate agent behavior, specifically utilizing commands like "STOP", "ANALYZE", and "REQUIRED Claude Behavior" to force a recursive follow-up loop. Additionally, the script scripts/ask_question.py automatically appends a "FOLLOW_UP_REMINDER" to all responses from NotebookLM, which contains instructions for the agent to review its own output and potentially execute more commands before finishing.
[PROMPT_INJECTION]: The "Smart Discovery" feature in SKILL.md creates a surface for indirect prompt injection. It instructs the agent to take content from a query (external data retrieved from NotebookLM) and use it directly as arguments for subsequent CLI commands (e.g., notebook_manager.py add). If the external content contains malicious characters or deceptive instructions, it could lead to unintended command execution or metadata poisoning.
[REMOTE_CODE_EXECUTION]: The initialization script scripts/init.py and the main runner scripts/run.py perform automated environment setup, which includes installing the patchright library and browser dependencies. This process involves executing setup scripts and binaries from the downloaded package at runtime.
[COMMAND_EXECUTION]: The skill makes extensive use of subprocess.run to manage virtual environments and execute browser automation setup. While intended for environment management, these are powerful system-level capabilities that interact with the host shell and represent a capability tier that could be abused if compromised.
[EXTERNAL_DOWNLOADS]: The skill automatically downloads and installs browser binaries (Chromium) during the first run via the patchright package manager. This involves fetching and executing code from external repositories outside of the skill's own package.

notebooklm