The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill features attack surfaces where untrusted data enters the agent context through URL fetching (WebFetch) and document reading (PDF/DOCX) in extract.md, reconcile.md, critique.md, and audit.md. Boundary markers are present in the form of explicit instructions to ignore embedded directives. The skill's capability inventory includes file writing and restricted Python execution. Sanitization is performed by the agent following instructions to discard behavioral suggestions and extract only facts. The risk is well-managed and consistent with the skill's primary purpose.
[REMOTE_CODE_EXECUTION]: The skill executes local Python scripts (slop_scan.py, validate_copy.py) via Bash. These scripts are contained within the skill, do not access the network, and use safe parsing libraries.
[DATA_EXFILTRATION]: No evidence of unauthorized data transmission was found. Network usage is initiated by the user for content extraction, and no sensitive system files are accessed.
[COMMAND_EXECUTION]: Command execution is limited to the intended Python-based validation tools provided with the skill.

copywriting