The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill fetches the pyhealth library from the official package registry and retrieves medical code resources and synthetic datasets from Google Cloud Storage (storage.googleapis.com/pyhealth/). These are vendor-related resources and well-known services used for research data distribution.
[DYNAMIC_EXECUTION]: The documentation identifies that the library's load_processors function utilizes the pickle module for deserializing state artifacts. The skill correctly includes warnings to only load trusted local processor artifacts to mitigate standard risks associated with untrusted data deserialization.
[INDIRECT_PROMPT_INJECTION]: The skill ingests external clinical data including EHR records (MIMIC-III/IV, eICU) and medical codes. While this represents a surface for indirect injection, the skill lacks boundary markers for the data content as it is intended for machine learning training rather than prompt-based control. The capabilities are restricted to model training and local file output for checkpoints.
Ingestion points: EHR records from MIMIC-III/IV and eICU datasets, medical code resources from remote CSV files.
Boundary markers: Not present for internal data records; handled by structured ML ingestion.
Capability inventory: Machine learning training, evaluation, and local checkpoint writing (Trainer.train, Trainer.evaluate, Trainer.load_ckpt).
Sanitization: Standard tokenization and numerical normalization provided by the library's dataset processors.
[COMMAND_EXECUTION]: The skill provides transparent commands to initialize research environments and run training pipelines using the uv tool, ensuring isolated execution and version-pinned dependency management.

pyhealth