The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests clinical data from external sources (such as MIMIC-III, MIMIC-IV, and eICU) to create supervised samples for model training, which constitutes an indirect prompt injection surface.\n
Ingestion points: Dataset root directories and cloud storage URLs defined in SKILL.md and references/datasets.md.\n
Boundary markers: Absent; the instructions do not specify markers to delimit untrusted data from the agent's context.\n
Capability inventory: The skill uses uv run and subprocess calls in Python scripts to train models and process data as described in SKILL.md and assets/starter_pipeline.py.\n
Sanitization: Absent; standard clinical data parsing is performed without specialized sanitization for injection patterns.\n- [EXTERNAL_DOWNLOADS]: The skill references clinical datasets and dependencies from well-known external services.\n
Evidence: Fetches synthetic data from https://storage.googleapis.com/pyhealth/Synthetic_MIMIC-III/ and specific PyTorch wheels from https://download.pytorch.org/whl/cu121.\n
Context: These sources are well-known services and organizations used for the library's legitimate operations.\n- [COMMAND_EXECUTION]: The skill provides instructions for the agent to execute shell commands to manage environments and run training tasks.\n
Evidence: uv add pyhealth, uv init, and uv run python train.py in SKILL.md and references/installation.md.\n
Context: These commands are necessary for the primary purpose of building and training healthcare ML models.

pyhealth