The Agent Skills Directory

[COMMAND_EXECUTION]: The scripts/run_eval.py script executes the claude CLI using subprocess.Popen to evaluate whether skill descriptions correctly trigger the agent. This execution uses list-based arguments to safely handle user-supplied queries, preventing shell injection.
[DYNAMIC_EXECUTION]: The eval-viewer/generate_review.py script starts a local web server (HTTPServer) on 127.0.0.1 to serve evaluation results and collect user feedback via a simple API. This is a functional requirement for the skill's visualization features.
[INDIRECT_PROMPT_INJECTION]: The skill processes user-provided evaluation queries and feedback, which are subsequently used in prompts for an automated description-optimization loop.
Ingestion points: Evaluation queries are read from eval_set.json in scripts/run_loop.py, and user feedback is read from feedback.json.
Boundary markers: Prompts in scripts/improve_description.py utilize XML-style tags (e.g., <current_description>, <scores_summary>, <new_description>) to delimit data components.
Capability inventory: The skill has the capability to execute the claude CLI via subprocess, write files to the workspace directory, and start a local HTTP server.
Sanitization: User-provided queries and feedback are wrapped in delimiters but do not undergo explicit sanitization before being interpolated into optimization prompts.

skill-creator