The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes external data from mobile app stores (App Store and Google Play) which represents a surface for instructions embedded in user reviews or descriptions to influence agent judgment.
Ingestion points: Store listings and reviews are retrieved via specialized sub-agents (copycat-scout, copycat-reviews) as defined in SKILL.md.
Boundary markers: The architecture uses independent 'Blue' and 'Red' sub-agents with restricted evidence contexts to provide separate judgment and validation stages.
Capability inventory: The skill manages local state via Python scripts and writes artifacts to the workspace directory.
Sanitization: The workflow relies on multi-agent verification and human checkpoints for selection and validation rather than simple string filtering.
[DYNAMIC_EXECUTION]: The skill uses a local CLI utility to maintain a deterministic run state and handle state transitions.
Evidence: scripts/copycat_run.py and the copycat_runtime module manage JSON state transitions using atomic file operations and include explicit path traversal protections when resolving run directories.
[UNVERIFIABLE_DEPENDENCIES_AND_REMOTE_CODE_EXECUTION]: The skill retrieves product metadata and customer evidence from well-known mobile application platforms for research purposes.
Evidence: scripts/copycat_store/google_protocol.py contains parsers for extracting data from Google Play's internal response formats.

copycat