The Agent Skills Directory

[PROMPT_INJECTION]: The skill instructions in SKILL.md and the output generated by ask_question.py use high-priority directives such as "Required Claude Behavior", "STOP", "ANALYZE", and "EXTREMELY IMPORTANT: Is that ALL you need to know?". These commands are designed to override the agent's natural decision-making and force a specific multi-step follow-up loop.
[COMMAND_EXECUTION]: The skill architecture relies on a universal wrapper (run.py) and setup scripts (setup_environment.py, __init__.py) that frequently invoke subprocess.run() to manage a local virtual environment, install dependencies, and execute other Python scripts within the skill directory.
[EXTERNAL_DOWNLOADS]: Upon first run, the skill automatically fetches and installs Python packages (patchright, python-dotenv) and browser binaries (Google Chrome) from external sources into a local .venv directory.
[INDIRECT_PROMPT_INJECTION]: The skill creates an attack surface by ingesting unstructured data from Google NotebookLM and presenting it to the agent. Furthermore, the ask_question.py script appends its own instructional context to the retrieved documentation, which acts as a multi-step prompt injection to influence subsequent agent actions.

notebooklm