The Agent Skills Directory

[COMMAND_EXECUTION]: The skill makes extensive use of the subprocess module to execute bundled scripts (e.g., run_loop.py, package_skill.py) and the claude CLI. These operations are core to the skill's functionality for creating, benchmarking, and packaging other skills and are executed without elevated privileges.
[EXTERNAL_DOWNLOADS]: The evaluation viewer components load the SheetJS library from a well-known public CDN (cdn.sheetjs.com). This is used to provide spreadsheet rendering capabilities for the user to review evaluation results.
[PROMPT_INJECTION]: The skill possesses a surface for indirect prompt injection because it ingests user-controlled test queries and feedback, which are then used as context for Claude to improve skill descriptions.
Ingestion points: Test queries from eval_set.json and user feedback from feedback.json enter the agent context.
Boundary markers: Prompt templates use XML-style tags like <scores_summary> and <skill_content> to separate untrusted data from instructions.
Capability inventory: The skill can write files and execute shell commands via the Claude CLI.
Sanitization: While untrusted content is wrapped in delimiters, the skill does not explicitly escape or sanitize the content to prevent potential delimiter breakouts.
[SAFE]: The code is well-structured and uses safe practices, such as yaml.safe_load() for parsing configuration files and providing human-in-the-loop review steps before finalizing skill modifications.

skill-creator