The Agent Skills Directory

[COMMAND_EXECUTION]: The converge.py script executes shell commands via subprocess.run to orchestrate existing skills (create-story, review-story). These calls are limited to pre-defined entry points (run.sh) and do not accept arbitrary user-controlled shell input. The implementation explicitly clears the VIRTUAL_ENV environment variable to prevent environment poisoning across skill executions.
[DATA_EXPOSURE]: The skill implements a structured convergence loop that tracks creative quality metrics (voice, lore, craft). While it interacts with a /memory skill, there is no evidence of sensitive credential harvesting or exfiltration to external domains. Access is restricted to relative paths within the skill directory structure.
[PROMPT_INJECTION]: The skill incorporates specific creative writing rules (Horus persona, Theory of Mind checks) and enforces that all critique must come from an automated subprocess (/review-story) rather than bespoke NLP. This architectural choice limits the surface area for direct or indirect prompt injection by decoupling evaluation from the generation prompt.

story-lab