The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill uses dynamic execution to load its own helper functions into the agent's environment. * Evidence: The setup instructions in SKILL.md recommend running exec(open("scvi-tools/kernel.py").read()).
[REMOTE_CODE_EXECUTION]: The skill provides instructions for delegating computation to external remote platforms. * Evidence: SKILL.md includes a code example for using the Modal serverless GPU platform (modal run pipeline.py).
[EXTERNAL_DOWNLOADS]: The skill requires the installation of several external Python dependencies from public registries. * Evidence: SKILL.md specifies installation of scvi-tools, scanpy, and anndata via pip.
[PROMPT_INJECTION]: The skill exhibits an attack surface for indirect prompt injection by processing external genomic data files. * Ingestion points: External .h5ad files are read using sc.read_h5ad in both SKILL.md and the provided pipeline.py script. * Boundary markers: There are no explicit markers or instructions provided to the agent to treat data content as non-instructional. * Capability inventory: The skill performs file system operations, dynamic script execution via exec, and network-based execution through external CLI tools. * Sanitization: Content from the data files is processed without validation or sanitization.

scvi-tools