The Agent Skills Directory

[PRIVILEGE_ESCALATION]: Documentation in references/r_interop.md instructs the agent to use sudo commands (e.g., sudo apt-get install, sudo dnf install) for installing system-level dependencies on Linux environments.
[DYNAMIC_EXECUTION]: The skill facilitates the generation and execution of scripts at runtime. For example, references/r_interop.md provides the complete source code for an R script (convert_rds_to_h5ad.R) and instructs the agent to save and execute it to handle data format conversions.
[EXTERNAL_DOWNLOADS]: The skill contains instructions for the runtime installation of several software packages from external repositories:
Python packages including scanpy[leiden], harmonypy, bbknn, and scikit-image are recommended for installation via uv pip.
R packages are fetched from cloud.r-project.org (a well-known service for R software distribution).
The seurat-disk package is installed directly from a GitHub repository (mojaveazure/seurat-disk) using the remotes R package.
[COMMAND_EXECUTION]: The CLI toolkit and analysis workflow depend on shell command execution for data processing, environment configuration, and running bundled Python scripts (e.g., run_pipeline.py, qc_analysis.py).
[INDIRECT_PROMPT_INJECTION]:
Ingestion points: Untrusted scientific data is ingested via load_anndata in scripts/_common.py and various sc.read_* calls across the toolkit, supporting H5AD, CSV, Loom, and MTX formats.
Boundary markers: No explicit markers or "ignore instructions" warnings are utilized during the processing of external data files.
Capability inventory: The skill possesses capabilities for local file system writes (adata.write_h5ad) and visualization generation. It also involves executing R and Python scripts based on processed inputs.
Sanitization: The skill does not perform sanitization on metadata fields (such as cell barcodes or gene symbols) which may be viewed or processed by the agent.

scanpy