The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes local commands such as the claude CLI and the lsof utility via Python's subprocess module. These calls facilitate running skill evaluations, optimizing trigger descriptions, and managing networking ports for the local viewer. Commands are executed using argument lists, which follows security best practices to prevent shell injection.
[EXTERNAL_DOWNLOADS]: The evaluation viewer (viewer.html) fetches the SheetJS library from its official CDN (cdn.sheetjs.com) to provide spreadsheet rendering capabilities within the local review interface. SheetJS is a well-known technology provider.
[DYNAMIC_EXECUTION]: The skill dynamically creates markdown command files within the project's .claude/commands/ directory to simulate skill availability during test runs. It also generates standalone HTML reports and temporary evaluation pages for performance tracking.
[DATA_EXPOSURE]: An HTTP server is launched on the local loopback interface (127.0.0.1) by generate_review.py to facilitate the qualitative review of evaluation results. Access is restricted to the local machine.
[INDIRECT_PROMPT_INJECTION]: The skill represents a surface for indirect prompt injection as it processes untrusted user-provided evaluation data and feedback.
Ingestion points: Evaluation queries are read from evals.json and user feedback is ingested via feedback.json into the agent's context.
Boundary markers: The skill does not use explicit boundary markers or delimiters when interpolating ingested queries into CLI arguments.
Capability inventory: The skill can execute shell commands (subprocess.run, Popen), write to the filesystem (including configuration directories), and start local network listeners.
Sanitization: The skill relies on the underlying platform's security guardrails and does not implement custom sanitization for ingested content before processing.

skill-creator