The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes unstructured text provided by the user via the $ARGUMENTS variable, which represents a surface for instructions embedded in data.
Ingestion points: Data enters the agent's context through the $ARGUMENTS parameter in the SKILL.md instruction block.
Boundary markers: The instructions do not specify explicit delimiters (like triple-backticks or unique separators) to encapsulate the user input.
Capability inventory: The skill has the ability to write to persistent storage using the nmem_remember tool and read from it using nmem_recall and nmem_context.
Sanitization: The skill implements a robust human-in-the-loop (HITL) control in Phase 5, which explicitly requires the user to review a preview of all gathered memories and provide a 'yes' confirmation before the nmem_remember tool is executed. This process serves as a critical defense against the unintended execution of malicious instructions found in the raw input.

memory-intake