The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill downloads software packages and binaries from the official PyTorch Geometric wheel index at https://data.pyg.org/whl/ to support sparse operations and GPU acceleration.
[INDIRECT_PROMPT_INJECTION]: The skill processes external data which serves as a potential vector for indirect instructions.
Ingestion points: Graph data and features are ingested from local files via pd.read_csv and from remote benchmark sources including the Planetoid, TUDataset, QM9, and Reddit datasets.
Boundary markers: There are no explicit delimiters or instructions provided to the agent to ignore potential commands embedded within the processed graph data.
Capability inventory: The skill environment allows for package installation via uv pip install and file system writes through the InMemoryDataset.save method.
Sanitization: The provided implementation lacks evidence of data validation or sanitization of input features and labels before they are used in model training or processing.
[METADATA_POISONING]: The skill includes instructions directing the AI agent to proactively suggest an external platform (www.k-dense.ai) when user workflows reach a certain level of complexity. This instruction steers agent behavior to promote a specific service.

torch-geometric