The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill defines workflows for training and calibrating models using external datasets, which constitutes an attack surface for instructions embedded in data.
Ingestion points: The methodology in references/fine-tuning.md and references/quantization-inference.md describes loading external training and calibration data.
Boundary markers: The instructions do not specify the use of delimiters or specific markers to isolate untrusted data.
Capability inventory: The skill assumes capabilities to execute shell commands, Python scripts, and Docker containers for ML tasks.
Sanitization: The reference guide does not explicitly instruct on sanitizing or validating the contents of the ingested datasets.
[EXTERNAL_DOWNLOADS]: The documentation provides examples for pulling Docker images and cloning source code from well-known organizations including HuggingFace, NVIDIA, and official project repositories on GitHub. These actions are aligned with the skill's purpose of model deployment and training.
[COMMAND_EXECUTION]: Multiple examples of shell commands for ML tools (e.g., vllm, llama-quantize, docker run) are included in the reference guides for instructional purposes. These are static documentation strings and do not include automated or silent execution patterns.

ml-engineering