The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill exhibits a significant attack surface for indirect prompt injection as it is designed to autonomously scan and process untrusted game project files.
Ingestion points: The skill reads extensive project data across multiple phases, including source code (*.cs, *.cpp, *.gd, *.js, *.ts), engine-specific configuration (ProjectSettings/, *.uproject, project.godot), and asset metadata (prefabs, scenes, localization files) as described in Phase 1 through Phase 8.
Boundary markers: The instructions lack any defined delimiters or explicit warnings for the agent to ignore instructions that might be embedded within the processed project files (e.g., within code comments or string data).
Capability inventory: The agent performs deep file inspection, generates detailed reports, and writes telemetry data. Furthermore, the "SELF-HEALING VALIDATION" section refers to "completing fixes," implying the agent has the capability to write or modify project files based on its analysis.
Sanitization: No sanitization or validation steps are defined for the external content before it is processed by the agent's logic.
[DATA_EXPOSURE]: The skill includes a "Self-Evolution Telemetry" mechanism that instructs the agent to write execution metadata to a hidden file at ~/.claude/projects/skill-telemetry.md. While this appears to be a platform-specific logging feature, writing data to hidden directories within the user's home folder is a sensitive behavior that could be exploited for unauthorized logging of project structures or user activity.

game-qa