The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill's primary function involves ingesting and processing external datasets and retrieved documents, creating a surface for indirect prompt injection where malicious instructions could be hidden in processed data.
Ingestion points: External data enters the system through scripts/analyze_dataset.py (CSV parsing), references/fine-tuning.md (JSONL dataset loading), and references/rag-patterns.md (retrieval of external document chunks).
Boundary markers: SKILL.md contains explicit guidance to treat external content as evidence rather than instructions. Prompt templates in references/rag-patterns.md include warnings for the model to ignore instructions embedded within the retrieved context.
Capability inventory: The skill enables model training, local file persistence for saving weights and metadata, and network operations for fetching pretrained models.
Sanitization: The instructions advocate for security best practices such as using trust_remote_code=False and pinning specific commit hashes for external dependencies.
[EXTERNAL_DOWNLOADS]: The skill facilitates the acquisition of pretrained models and datasets from established machine learning registries.
Evidence: Code patterns in references/fine-tuning.md and references/transformers-patterns.md download artifacts from HuggingFace (e.g., meta-llama/Llama-2-7b-hf).
[COMMAND_EXECUTION]: The skill identifies and provides patterns for executing standard machine learning command-line utilities.
Evidence: Documentation covers the use of tools such as vllm for serving, optimum-cli for ONNX model export, and torchserve for deployment.

engineering-ml-engineer