The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from external .h5ad (AnnData) files, which could potentially contain malicious instructions embedded in metadata or labels designed to influence the agent's behavior.
Ingestion points: cells.h5ad as defined in SKILL.md and references/scgpt_usage.md.
Boundary markers: No specific delimiters or instructions to ignore potential instructions within the data files are present in the skill body.
Capability inventory: The skill is granted Read, Write, and Edit permissions, and can execute shell commands via Bash(python:*) and Bash(uv:*) to process data and run models.
Sanitization: No data sanitization or schema validation steps are described for the input files.
[EXTERNAL_DOWNLOADS]: The skill references and downloads large pretrained model checkpoints from the bowang-lab/scGPT repository and associated infrastructure. This is a standard and necessary component for the functionality of a foundation model skill.
[COMMAND_EXECUTION]: The skill utilizes uv run python and standard Python interpreters to execute scientific code. While these are necessary for the skill's purpose, they represent a capability to run complex logic based on the user's data and the skill's instructions.

alterlab-scgpt