The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/mlx_smoke.py uses subprocess.run to invoke the mlx_lm training module. This is the primary function of the skill, used to perform local model fine-tuning on Apple Silicon.
[INDIRECT_PROMPT_INJECTION]: The skill ingests training data from external sources (Busabase and Kelly Support). While this data is processed by the agent, the skill implements multiple security controls:
Mandatory human review and approval steps in the workflow (approve, request_changes, block).
Strict schema validation for app_spec tasks to ensure data follows expected JSON structures.
Deterministic evaluation using locked test datasets to measure model performance accurately.
[EXTERNAL_DOWNLOADS]: The skill manages dependencies via standard package managers and fetches model weights from Hugging Face through the mlx-lm library, which is consistent with its purpose as a model training lab.

kelly-local-model-lab