The Agent Skills Directory

[COMMAND_EXECUTION]: The skill includes a local helper script scripts/contrast.py used for anonymizing candidate identifiers and validating that lessons contain empirical evidence. This script is written in pure Python using standard libraries and operates entirely offline.
[INDIRECT_PROMPT_INJECTION]: The refinement process involves the agent reading and comparing outputs generated by other skill versions. This creates an attack surface where instructions embedded in those outputs could influence the agent. However, the skill's explicit requirement for blind selection and evidence-based composition acts as a structural defense against unverified behavioral changes.

contrastive-refinement