The Agent Skills Directory

[COMMAND_EXECUTION]: Several scripts, including scripts/run_eval.py and scripts/improve_description.py, execute the claude CLI via subprocesses to test skill triggers and generate improvements. These calls are properly parameterized and do not use a shell, mitigating command injection risks.
[COMMAND_EXECUTION]: The eval-viewer/generate_review.py script uses system utilities like lsof and kill to manage the local server's port. These calls are restricted to specific, safe arguments.
[PROMPT_INJECTION]: The skill has an indirect prompt injection surface because it interpolates untrusted user data—such as test queries and evaluation feedback—directly into prompts for subagents. While this is a vulnerability surface, the risk is localized to the development cycle of new skills.
[DATA_EXFILTRATION]: The eval-viewer/generate_review.py script starts a local HTTP server binding to 127.0.0.1. This server reads and serves files (including base64-encoded PDF and XLSX files) from the evaluation workspace to the user's browser for review.
[SAFE]: The skill implements best practices for handling data, such as using yaml.safe_load in scripts/quick_validate.py to prevent unsafe deserialization of skill metadata.

skill-creator