The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes the claude CLI via subprocess calls in scripts/run_eval.py and scripts/improve_description.py. This is the primary method for performing automated evaluations and description optimization within the target environment.
[EXTERNAL_DOWNLOADS]: The evaluation viewer (eval-viewer/viewer.html) loads the SheetJS library from a well-known CDN (cdn.sheetjs.com) to render spreadsheet files inline. This is a legitimate use of a common library for data visualization.
[INDIRECT_PROMPT_INJECTION]: The skill ingests and processes user-provided test cases, scripts, and evaluation outputs. While this creates a surface for indirect injection, the scripts utilize HTML escaping when generating reports, and the execution occurs within the controlled testing loop of the development environment.
Ingestion points: Reads evals/evals.json, feedback.json, and files within the specified workspace outputs/ directories.
Boundary markers: The grader.md instructions guide the agent to evaluate evidence from outputs but do not enforce technical delimiters.
Capability inventory: Includes file system writes (init_skill.py), local CLI execution of claude (run_eval.py, improve_description.py), and local HTTP server hosting (generate_review.py).
Sanitization: generate_review.py and generate_report.py use HTML escaping when embedding content into the viewer.
[COMMAND_EXECUTION]: scripts/init_skill.py uses chmod to set executable permissions on generated scripts, which is standard for initializing a development project.
[COMMAND_EXECUTION]: eval-viewer/generate_review.py starts a local HTTP server on a user-defined port (default 3117) to serve the evaluation review interface, restricted to 127.0.0.1.

skill-creator