The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted user input by saving it to a local workspace (_workspace/{run_id}/01_input.txt) and passing it through multiple sub-agents (e.g., humanize-monolith, korean-style-rewriter). This creates a vulnerability where malicious instructions embedded in the input text could attempt to override the sub-agents' behavior.
Ingestion points: The skill accepts user-supplied Korean text for processing and saves it to a temporary workspace file.
Boundary markers: The instructions prioritize 'fidelity' and 'meaning invariance' but do not explicitly mandate the use of robust delimiters (such as XML tags) when interpolating user data into the prompts for sub-agents.
Capability inventory: The skill has the ability to read and write files within its workspace and call other agents using the Agent tool.
Sanitization: The skill uses a 'fidelity auditor' and 'naturalness reviewer' to validate the output against the original meaning, providing a level of post-processing verification, though no pre-processing sanitization of input is performed.

humanize-korean