The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes the kaggle command-line tool via the subprocess module in scripts/submission_quota.py and scripts/upload_dataset.py. These executions are parameterized using argument lists rather than shell strings, which prevents shell-level injection. The capabilities are aligned with the skill's stated purpose and permissions.\n- [PROMPT_INJECTION]: The skill has an indirect prompt injection surface because it processes untrusted data from Kaggle (e.g., in fetch_competition_info.py, discussion_ingest.py, and kernel_read.py). \n
Ingestion points: Data enters via API and CLI calls fetching competition details, forum posts, and kernel source.\n
Boundary markers: The agent renders content into markdown sections, though explicit instructions to ignore embedded commands within the fetched data are primarily handled by the LLM's own safety guardrails.\n
Capability inventory: The agent has the ability to run shell commands (Kaggle CLI), perform network operations, and write to the local filesystem.\n
Sanitization: Content is cleaned via html_to_markdown and sanitize_cli_output, which strips control characters and ANSI escape sequences to prevent terminal spoofing.\n- [EXTERNAL_DOWNLOADS]: During initialization, the skill downloads several official and well-known Python packages from PyPI (httpx, kaggle, kagglesdk, nbformat, pydantic, python-dotenv, rich). These are required for interaction with the Kaggle platform and are treated as trusted dependencies.

nvidia-kaggle-skill