The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill documents workflows that ingest data from external sources and process them through Large Language Models (LLMs).
Ingestion points: The rag_ingestion_pipeline fetches documents from user-provided source_url strings (SKILL.md), and the batch_inference_flow reads data from an input_file (SKILL.md).
Boundary markers: The provided design sketches do not include specific delimiters or instructions for the model to ignore embedded commands within the processed data.
Capability inventory: The pipelines have access to network operations (fetching documents), file system writes (storing results), and cloud compute resources via the KubernetesPodOperator (Airflow) or the OpenAI API (Prefect).
Sanitization: No sanitization or validation of the fetched document content or input file lines is implemented before they are chunked, embedded, or sent to the LLM.
[EXTERNAL_DOWNLOADS]: The documentation includes instructions to install third-party packages from public registries.
Evidence: The skill suggests running pip install prefect prefect-kubernetes and imports libraries such as sentence-transformers, qdrant-client, and openai (SKILL.md). These are well-known industry-standard libraries used for AI and orchestration tasks.

ai-pipeline-orchestration