The Agent Skills Directory

[COMMAND_EXECUTION]: The skill automates the testing of AI behaviors by invoking the claude CLI tool through Python's subprocess module. These operations are restricted to non-shell executions and are core to the skill's function of benchmarking performance.
[INDIRECT_PROMPT_INJECTION]: The skill ingests and processes untrusted user data, including task prompts and feedback, which are stored in JSON format and rendered in a local web viewer (viewer.html). The skill includes explicit instructions to the AI agent to avoid generating malicious or misleading content, mitigating the risk of providing harmful outputs to the user.
[REMOTE_CODE_EXECUTION]: The improve_description.py script makes outbound network calls to the Anthropic API to optimize skill metadata. This is a documented and expected behavior for a tool designed to interface with large language models.
[FILE_SYSTEM_ACCESS]: The skill manages files within the user's workspace, specifically in the .claude/commands/ and evals/ directories, to facilitate the registration and testing of new skills. This behavior is transparent and aligned with the tool's primary purpose.

skill-creator