The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill possesses a data ingestion surface by reading external script files (scripts/<id>.md) and rubric definition files. However, it implements a security-first 'Channel B' architecture designed to prevent prompt injection and data contamination. Key defenses include:
Ingestion Points: Limited strictly to <script-path> and <rubric-notes-path> passed via task tool.
Boundary Markers: Explicit 'Hard Refusal' lists that prevent reading of sensitive files like .cheat-state.json, predictions/, and videos/ which could leak performance data.
Capabilities: Uses Read, Glob, and Grep for file inspection within a defined scope.
Sanitization: Implements an automated self-check using Grep to detect real-world performance metrics (e.g., play counts, likes) in input files, which triggers a refusal or a confidence reduction.
[COMMAND_EXECUTION]: The skill utilizes platform tools (Read, Glob, Grep) to evaluate content. These tools are used for legitimate local file system operations necessary for the skill's primary function and are limited to the project workspace.
[SAFE]: The skill demonstrates safe practices, such as requiring structured JSON output and prohibiting the use of markdown explanations to prevent downstream parsing errors. No credential theft, obfuscation, or unauthorized network operations were detected.

cheat-score-blind