The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill instructs the agent to use the trust_remote_code=True flag when loading specific scientific models from Hugging Face repositories. This flag allows the execution of arbitrary Python code contained within the remote model repository during the loading process. While this is a standard requirement for many custom architectures in the scientific ML community, it constitutes a remote code execution vector if a model repository is compromised. Evidence is found in references/using-models.md and SKILL.md.
[PROMPT_INJECTION]: The skill fetches content from an external catalog at huggingscience.co and parses it to guide model and dataset selection, introducing a surface for indirect prompt injection.
Ingestion points: The scripts/fetch_catalog.py script fetches data from huggingscience.co/llms-full.txt and domain-specific markdown files.
Boundary markers: Absent; the content is parsed into structured objects but presented to the agent without explicit delimiters or instructions to ignore embedded commands.
Capability inventory: The agent has capabilities for shell execution (uv run), file writing (via .env creation instructions), and network operations (transformers, gradio_client).
Sanitization: Absent; the description and tags from the remote catalog are used directly without validation.
[EXTERNAL_DOWNLOADS]: The skill performs automated network requests to huggingscience.co to retrieve catalog indices and resource descriptions. These downloads are central to the skill's discovery functionality but represent a dependency on external infrastructure.
[COMMAND_EXECUTION]: The skill instructions routinely direct the agent to execute the bundled Python script scripts/fetch_catalog.py via uv run to discover and filter scientific resources. This is an expected behavior for the skill's functionality.

hugging-science