The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill instructions and documentation recommend the installation of standard bioinformatics packages from well-known official registries such as PyPI and Bioconductor.
Evidence: pip install cell2location and pip install tangram-sc in usage-guide.md.
Evidence: Recommendations to install Bioconductor packages spacexr, SPOTlight, and STdeconvolve.
[INDIRECT_PROMPT_INJECTION]: The skill involves the ingestion and analysis of external transcriptomics data files (Anndata/H5AD format) where the agent is tasked with interpreting labels and determining the resolution regime. This introduces an attack surface for indirect prompt injection via malicious metadata in user-provided files.
Ingestion points: The agent reads reference_scrna.h5ad and visium.h5ad into the execution environment in SKILL.md and examples/deconvolve_spatial.py.
Boundary markers: There are no explicit instructions or delimiters provided to the agent to treat external file content as untrusted data or to ignore embedded instructions.
Capability inventory: The skill possesses filesystem capabilities, including writing results (adata_vis.write_h5ad) and deleting temporary directories (shutil.rmtree).
Sanitization: No sanitization or validation logic is applied to the metadata or cell labels extracted from the datasets before they are used to inform agent conclusions.

bio-spatial-transcriptomics-spatial-deconvolution