prompt-engineering-patterns
Pass
Audited by Gen Agent Trust Hub on Sep 14, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill implements templates designed to process untrusted external data (e.g., RAG contexts, user queries, and article summaries).
- Ingestion points: Untrusted data enters the agent context through variables such as
{text},{query},{context}, and{problem}acrossSKILL.md,assets/prompt-template-library.md, andreferences/chain-of-thought.md. - Boundary markers: The templates frequently incorporate explicit instructions to the model to restrict its scope (e.g., "Answer ONLY based on the provided context" in
RAG_PROMPT) and use structural delimiters for clarity. - Capability inventory: The included utility script
scripts/optimize-prompt.pyand references use standard libraries (numpy,langchain,pydantic) andThreadPoolExecutorfor parallel evaluation; no dangerous system commands, persistent access, or exfiltration capabilities are identified. - Sanitization: The skill promotes security best practices by teaching chain-of-thought verification, self-consistency checks, and schema enforcement using Pydantic to ensure model outputs are reliable and structured.
Audit Metadata