The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes the subprocess module to execute system commands. Specifically, scripts/run_eval.py invokes the claude CLI for evaluations, and eval-viewer/generate_review.py runs lsof to manage local network ports.
[COMMAND_EXECUTION]: The script eval-viewer/generate_review.py starts a local HTTP server using the HTTPServer class to host an interactive evaluation results dashboard.
[COMMAND_EXECUTION]: The skill performs dynamic code generation in scripts/run_eval.py by writing markdown-based command descriptors to the .claude/commands/ directory, which are then loaded and executed by the CLI environment.
[REMOTE_CODE_EXECUTION]: The skill utilizes the official anthropic SDK to communicate with Anthropic's API for the description optimization workflow.
[EXTERNAL_DOWNLOADS]: The evaluation viewer (eval-viewer/viewer.html) fetches the xlsx.full.min.js library from cdn.sheetjs.com, a well-known service, at runtime to support browser-based spreadsheet rendering.
[PROMPT_INJECTION]: The skill exhibits an indirect prompt injection surface by processing untrusted data as core input for its workflows.
Ingestion points: Untrusted data enters the agent context via evals/evals.json (test prompts) and feedback.json (user reviews).
Boundary markers: The skill does not implement delimiters or 'ignore' instructions when interpolating this data into CLI queries or optimization prompts.
Capability inventory: The skill has access to shell execution (subprocess), filesystem writes, and network operations.
Sanitization: There is no evidence of validation or sanitization of external content before it is used to influence the agent's behavior or optimize skill descriptions.

skill-creator