The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill implements a workflow where an agent reads code files generated in a previous phase and applies self-correction logic based on its findings. This creates an attack surface where malicious instructions hidden in the code being analyzed (e.g., within comments or strings) could potentially override the skill's instructions during the analysis or correction phases.
Ingestion points: The skill reads all modified files defined in the {code_files} list during step-01-analyze-code.md.
Boundary markers: There are no defined delimiters or specific instructions for the agent to treat the file content as purely passive data or to ignore any natural language instructions found within those files.
Capability inventory: The skill grants the agent the ability to write to and modify files in step-04-self-correct.md and to persist JSON metadata in the project directory in step-05-report.md.
Sanitization: The skill does not perform any escaping, filtering, or validation of the code content before the agent processes it to determine necessary corrections.

elicitation