The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes local commands using Python's subprocess module in scripts like run_eval.py, improve_description.py, and generate_review.py. These calls are used to invoke the claude CLI for testing, manage background processes, and check system status (e.g., using lsof to manage the viewer's port). This is a standard pattern for the skill's documented purpose.
[EXTERNAL_DOWNLOADS]: The viewer.html file includes a reference to the SheetJS library via a public CDN (cdn.sheetjs.com). This is used to render spreadsheet files within the evaluation viewer and is a well-known service for this purpose.
[INDIRECT_PROMPT_INJECTION]: The skill acts as an orchestration layer that processes untrusted inputs from user-defined evals.json and feedback files.
Ingestion points: scripts/run_eval.py reads user task prompts from the eval set; scripts/run_loop.py and scripts/improve_description.py process evaluation results and feedback to refine descriptions.
Boundary markers: The skill uses XML-style tags (e.g., <skill_content>, <new_description>) to isolate untrusted content when communicating with the agent during the improvement loop.
Capability inventory: The skill can execute shell commands, perform local file operations, and manage network ports (localhost only).
Sanitization: While the skill primarily executes the provided prompts for testing purposes, it employs structured delimiters to maintain context separation and prevent unintended instruction following.

skill-creator