The Agent Skills Directory

[COMMAND_EXECUTION]: The scripts run_eval.py, run_task_eval.py, and improve_description.py make extensive use of subprocess.Popen and subprocess.run to execute the copilot CLI tool. These calls incorporate user-provided queries and prompts directly into the command arguments.
[REMOTE_CODE_EXECUTION]: The skill implements a workflow that dynamically generates SKILL.md files and temporary Python scripts (such as pivot_sales.py), which are then executed by the agent. This allows for the execution of arbitrary instructions generated or modified during the skill creation process.
[DATA_EXFILTRATION]: The file pivot_sales.py is hardcoded to read from and write to the user's home directory (Path.home() / "data"). While intended for processing sales data, this pattern demonstrates the capability to access sensitive user files.
[PROMPT_INJECTION]: The skill processes untrusted user data (eval queries and qualitative feedback) and interpolates them into agent prompts and subprocess execution lists. This presents an indirect prompt injection surface as documented in SKILL.md and handled in run_eval.py.
[EXTERNAL_DOWNLOADS]: The viewer.html template loads the xlsx library from an external CDN (cdn.sheetjs.com) at runtime to render spreadsheet data in the browser.

skill-creator