The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill provides templates and architectural patterns for RAG (Retrieval-Augmented Generation) and inference APIs that ingest and process untrusted external data.
Ingestion points: nlp.md defines a RAGSystem class, and mlops.md provides a FastAPI PredictionRequest schema for accepting user features.
Boundary markers: The skill includes explicit guidelines in SKILL.md and evaluation-and-guardrails.md for implementing prompt injection resistance, refusal policies, and output validation.
Capability inventory: The templates include model execution, local file reading (model artifacts), and network operations for logging and monitoring.
Sanitization: nlp.md provides a clean_text utility that performs Unicode normalization (NFKC) and removes potentially problematic characters.
[DYNAMIC_EXECUTION]: The skill references standard machine learning deserialization practices necessary for model serving.
mlops.md includes templates using joblib.load to restore scikit-learn model artifacts.
transformers-llm.md and deep-learning.md provide code for loading models via torch.load and torch.jit.load.
These operations are routine for AI/ML workflows and focus on loading artifacts from a governed model registry.
[EXTERNAL_DOWNLOADS]: The skill references and integrates with well-known, trusted AI/ML registries and platforms.
Fetches models and tokenizers from HuggingFace using the transformers library (AutoModel.from_pretrained).
Integrates with industry-standard MLOps services like Weights & Biases and MLflow for experiment tracking.
References official container images from NVIDIA and Python for training and serving environments.
[COMMAND_EXECUTION]: Provides standard developer workflows for validation and automated deployment.
SKILL.md lists benign validation commands such as pytest, ruff, and mypy.
mlops.md includes standard GitHub Actions workflows and Docker configuration for continuous integration and deployment.

ai-ml-principal-engineer