The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted single-cell data loaded from various file formats (.h5ad, .csv, Visium), which serves as an ingestion point for data that could contain malicious instructions embedded in metadata or feature names.
Ingestion points: The skill frequently uses data loading functions such as sc.read_h5ad, sc.read_csv, sc.read_10x_mtx, and sc.read_visium across its reference documentation.
Boundary markers: No specific delimiting instructions or warnings to ignore embedded content within ingested data are provided.
Capability inventory: The skill utilizes capabilities for file system writes (model.save, adata.write, de_results.to_csv) and performs network fetches (add_dna_sequence, arXiv metadata fetch).
Sanitization: There is no evidence of data validation or sanitization of observation metadata or feature names before they are processed by the agent.
[EXTERNAL_DOWNLOADS]: The skill provides procedures for downloading genomic references and scientific datasets from well-known repositories.
The function scvi.data.add_dna_sequence is used with install_genome=True to download large genomic data files (e.g., hg38).
Tutorial datasets are fetched via scvi.data.heart_cell_atlas_subsampled.
The skill documentation describes fetching publication metadata from export.arxiv.org for citation formatting.
[COMMAND_EXECUTION]: The skill instructs the user or agent to execute shell commands for environment preparation and dependency management.
Standard commands like uv pip install scvi-tools and uv pip install "scvi-tools[cuda]" are used for installation.

scvi-tools