The Agent Skills Directory

[COMMAND_EXECUTION]: The skill's bundled Python scripts utilize the subprocess module to programmatically call the claude CLI and various local utility scripts (aggregate_benchmark, run_loop, package_skill) to automate the skill-development lifecycle.
[PROMPT_INJECTION]: The skill processes untrusted user data (test queries) which are passed to subagents for execution, creating an indirect prompt injection surface.
Ingestion points: Untrusted data enters the agent context through the evals/evals.json configuration file and user-provided conversation history.
Boundary markers: The run_eval.py script uses YAML block scalars to encapsulate skill descriptions in temporary command files but lacks explicit delimiters to prevent subagents from obeying instructions embedded within test queries.
Capability inventory: The skill possesses capabilities to execute shell commands, perform filesystem read/write operations (including within the sensitive .claude/commands/ directory), and host a local HTTP server.
Sanitization: The skill performs minimal sanitization or escaping of external test queries before they are interpolated into prompts and passed to the claude CLI.
[EXTERNAL_DOWNLOADS]: The HTML review viewer references the SheetJS library and Google Fonts via well-known Content Delivery Networks (CDNs) to support data visualization and typography. These resources are from established providers and are used for their intended functional purposes.

skill-creator