The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill's setup process installs the Google Chrome browser via the patchright library to facilitate browser automation. This is a functional requirement for interacting with the NotebookLM web interface.
[COMMAND_EXECUTION]: The skill uses subprocess.run to manage its internal environment, install dependencies from a locked requirements file, and execute its own management scripts. These operations are restricted to the skill's local directory and managed virtual environment.
[INDIRECT_PROMPT_INJECTION]: The skill processes data from external notebooks which could contain malicious instructions designed to influence the agent's behavior.
Ingestion points: Content is scraped from the Google Notebook web interface in ask_question.py and browser_session.py.
Boundary markers: The skill explicitly identifies answers as "untrusted text" and provides clear instructions to the agent to treat the content as data and ignore any embedded instructions.
Capability inventory: The skill has permissions for network access (to Google domains) and local file writes (for library management and session state).
Sanitization: The skill separates retrieved data from the agent's control logic using distinct response fields and reminder strings (FOLLOW_UP_REMINDER).
[CREDENTIALS_SAFE]: The skill handles Google authentication through a manual login process in a visible browser. Session tokens and cookies are stored locally in the data/ directory. Access is secured using POSIX-style restrictive permissions (0o600 for files and 0o700 for directories) to prevent unauthorized local access.

blog-notebooklm