The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes the 'claude' CLI tool to perform evaluations and optimize descriptions. These calls are performed using the subprocess module with arguments passed as a list, which prevents shell injection.
Evidence: Found in scripts/improve_description.py (line 62) and scripts/run_eval.py (line 78).
[COMMAND_EXECUTION]: The results viewer script executes the 'lsof' command to manage its local server's port allocation.
Evidence: Found in eval-viewer/generate_review.py (line 197).
[DATA_EXFILTRATION]: The skill starts a local HTTP server on '127.0.0.1' to allow the user to review evaluation results and feedback. This server reads files from the designated workspace directory to render them in the viewer.
Evidence: Found in eval-viewer/generate_review.py (line 228) and the accompanying viewer.html.
[PROMPT_INJECTION]: The skill exhibits an indirect prompt injection surface by interpolating user-provided skill content and evaluation queries into prompts used for description optimization. It mitigates this risk by wrapping external content in XML-like tags (e.g., <skill_content>, <scores_summary>) to provide clear boundaries for the model.
Ingestion points: scripts/improve_description.py reads data from eval_results and skill_content.
Boundary markers: Present in the improve_description function's prompt template.
Capability inventory: Spawns subagents, writes files to .claude/commands/, and executes local CLI commands.
Sanitization: Uses structured XML-style delimiters to separate untrusted data from instructions.
[SAFE]: The skill uses 'cdn.sheetjs.com' to load the SheetJS library for rendering Excel files in the results viewer. This is a well-known service for client-side spreadsheet processing.
Evidence: Found in eval-viewer/viewer.html (line 8).

skill-creator