The Agent Skills Directory

[PROMPT_INJECTION]: The skill defines a workflow for processing a user-provided --target file or directory. This creates an indirect prompt injection surface where malicious instructions within a processed file could influence the agent's behavior during the OBSERVE, THINK, or EXECUTE phases of the improvement cycle.
Ingestion points: Untrusted content enters the agent's context through the --target file path specified in the /loop command (SKILL.md).
Capability inventory: The iterative algorithm includes a BUILD and EXECUTE phase, which provides a high-privilege context for potential instructions hidden in the target data (SKILL.md).
Boundary markers: The skill does not specify any delimiters or instructions to the agent to disregard embedded directives within the target content.
Sanitization: No sanitization or validation of the input target file is described in the workflow.

Loop