The Agent Skills Directory

[PROMPT_INJECTION]: The skill uses instructional prompts with high-emphasis markers intended to override or strongly influence the agent's conversational behavior.
Evidence: In scripts/ask_question.py, the constant FOLLOW_UP_REMINDER is appended to answers and starts with "EXTREMELY IMPORTANT:", instructing the AI to check for clarity and ask follow-up questions before replying.
[COMMAND_EXECUTION]: The skill utilizes automated browser interactions to perform tasks on behalf of the user.
Evidence: The scripts scripts/ask_question.py and scripts/remote_manager.py use the Playwright library to navigate the Google NotebookLM interface, fill forms, and simulate user input to automate complex workflows.
[DATA_EXPOSURE]: The skill manages sensitive session data to provide persistent authentication.
Evidence: scripts/auth_manager.py extracts and validates critical Google session cookies (such as SID, HSID, and SSID) and stores them in persistent browser profiles within the user's home directory at ~/.config/claude/notebooklm-skill/.
[INDIRECT_PROMPT_INJECTION]: The skill retrieves data from a remote web service, which could contain instructions that influence the agent's logic.
Ingestion points: scripts/ask_question.py extracts text directly from the NotebookLM web UI via CSS selectors in _collect_response_texts.
Boundary markers: No delimiters or boundary markers are used to separate the fetched content from the agent's internal instructions.
Capability inventory: The skill can read local files, write files, and perform network requests via Playwright.
Sanitization: External content is not sanitized or validated before being returned to the AI agent.

notebooklm