The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill provides implementation templates for Retrieval-Augmented Generation (RAG) pipelines in SKILL.md and references/rag-pipelines.md, which creates an architectural surface for indirect prompt injection.
Ingestion points: External text data is ingested from vector databases (such as LanceDB or pgvector) through the retrieve method.
Boundary markers: The context assembly logic in references/rag-pipelines.md utilizes simple triple-dash (---) delimiters to separate retrieved document chunks but lacks robust markers or warnings to prevent the LLM from obeying instructions embedded within the retrieved data.
Capability inventory: The skill possesses the capability to read from databases and interpolate that content directly into LLM prompts using the openai client, which then generates a response for the user.
Sanitization: The implementation patterns do not include explicit sanitization or filtering of the retrieved external content before it is processed by the LLM.
[EXTERNAL_DOWNLOADS]: The skill recommends and provides code for installing standard third-party libraries for AI development. These include well-known and trusted packages such as openai, polars, lancedb, duckdb, sentence-transformers, tenacity, and langchain.
[DATA_EXFILTRATION]: The LLMMonitor class in references/monitoring.md includes functionality to log full prompts and responses to a local DuckDB file for observability purposes. While the skill documentation identifies this as a potential privacy concern and defaults to an opt-out configuration (log_full_prompts=False), the storage of raw interaction data in a local database creates a data exposure surface that requires careful credential and file-system management.

engineering-ai-pipelines