The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from multiple sources including URLs, YouTube links, Google Docs, and local files, which are subsequently processed by the NotebookLM interface.\n
Ingestion points: SKILL.md defines sub-flows for adding sources (URL, text, file, Google Doc) and querying notebooks via chat interface.\n
Boundary markers: The skill uses structured prompt templates in scripts/custom_prompt_template_generator.py but lacks explicit instructions for the target LLM to ignore instructions embedded within the ingested source material.\n
Capability inventory: The skill utilizes browser automation capabilities (navigate, click, type, screenshot) and a dedicated file upload tool.\n
Sanitization: External content and user-provided parameters are interpolated into prompt templates and typed into the browser without explicit filtering or escaping.\n- [COMMAND_EXECUTION]: The skill accepts absolute file paths from the user to perform file uploads through the browser automation environment and executes local scripts for action routing and classification.\n
Evidence: SKILL.md specifies a sub-flow for 'File Upload' using user-provided absolute paths and invokes scripts such as scripts/async_action_classifier.py for operational logic.\n- [EXTERNAL_DOWNLOADS]: The skill navigates to Google's NotebookLM service and may fetch content from external websites as directed by the user.\n
Evidence: SKILL.md directs the agent to navigate to https://notebooklm.google.com and provides functionality to add sources via URLs. These actions target well-known services or are user-directed, and are documented as part of the skill's primary function.

notebooklm