The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill implements multiple patterns for ingesting untrusted external data that could contain malicious instructions designed to influence downstream LLM tasks.
Ingestion points: The fetch_documents task in the Prefect RAG pipeline accepts a source_url for external content, and the batch_inference_flow reads from a user-provided input_file.
Boundary markers: The provided code snippets do not include explicit delimiters or "ignore embedded instructions" warnings for the ingested content.
Capability inventory: The ingested data is processed via SentenceTransformer for embeddings and passed to OpenAI's gpt-4o-mini in the batch inference pipeline.
Sanitization: No explicit sanitization or validation of the fetched document content or input file data is shown before processing.
[COMMAND_EXECUTION]: The skill provides instructions and code for executing commands and managing infrastructure.
The setup instructions include local shell commands for package installation and starting orchestration servers.
The Airflow example utilizes the KubernetesPodOperator to execute the axolotl training CLI within a remote container (nvcr.io/nvidia/pytorch:24.05-py3), which is a standard pattern for model training orchestration.
[EXTERNAL_DOWNLOADS]: The skill references several external resources and libraries.
Downloads and installs standard Python packages including prefect, prefect-kubernetes, sentence-transformers, qdrant-client, and openai.
Fetches the official nvidia/pytorch container image from NVIDIA's registry (nvcr.io), which is a well-known service.

ai-pipeline-orchestration