The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data from CubeCobra (via IDs/URLs) or local files (CSV, JSON, plain text) which may contain adversarial instructions embedded in metadata or description fields. This content is used to brief the agent during Phase 2, Step 6 (Self-Grill), creating a surface where malicious input could influence the agent's tuning recommendations.
Ingestion points: cubecobra-fetch utility output and local file parsing in parse-cube (specifically the designer_intent.description and stated_archetypes fields).
Boundary markers: No specific delimiters or warnings to ignore embedded instructions are implemented for the ingested free-text fields.
Capability inventory: The skill utilizes shell command execution for MTG analysis tools (e.g., scryfall-lookup, cube-balance, archetype-audit) and performs file system writes via the Write tool.
Sanitization: No sanitization, escaping, or validation logic is specified for the external free-text data before it is interpolated into the agent's context.

cube-wizard