The Agent Skills Directory

Indirect Prompt Injection (MEDIUM): The skill provides integration patterns for RAG and template systems that are vulnerable to injection attacks.
Ingestion points: Untrusted data enters via variables such as {retrieved_context}, {user_question}, and {query} in SKILL.md (Integration Patterns and Quick Start sections).
Boundary markers: Absent. The templates use simple string interpolation (f-strings) without delimiters or 'ignore embedded instructions' headers, which can lead the model to follow instructions hidden in the retrieved context.
Capability inventory: The skill references a local execution capability via scripts/optimize-prompt.py and the prompt_optimizer library.
Sanitization: No evidence of sanitization or validation for the interpolated strings is provided in the examples.
Unverifiable Dependencies (MEDIUM): The code snippets rely on a non-standard library prompt_optimizer and reference a local script scripts/optimize-prompt.py. If these are not vetted, they represent a risk of executing unverified logic during prompt construction or optimization.

prompt-engineering-patterns