The Agent Skills Directory

[PROMPT_INJECTION]: The skill implements a self-optimization loop where it modifies its own core prompt templates based on historical performance and quality signals. This creates a surface for indirect prompt injection, as adversarial data processed by the agent could influence the optimizer to adopt and promote malicious prompt variations.
Ingestion points: Reads performance data and prompt logs from files in the workspace/memory/prompts/ directory.
Boundary markers: The skill description does not specify the use of delimiters or sanitization when incorporating successful variations back into primary prompt templates.
Capability inventory: The skill has file:read and file:write permissions to access and update its internal memory and prompt configurations.
Sanitization: There is no mention of filtering or sanitizing the generated prompt variations to prevent the inclusion of injected instructions.

prompt-optimizer