The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes the subprocess module across several scripts (run_eval.py, improve_description.py, generate_review.py) to orchestrate tasks via the claude CLI and manage local processes for its internal web server. This execution is consistent with the skill's purpose as a developer tool and is used to automate command-line workflows.
[DYNAMIC_EXECUTION]: The skill dynamically generates temporary command definitions and skill metadata in the .claude/commands/ directory to simulate triggering during evaluation runs. It also delegates specialized tasks to subagents (grader, comparator, analyzer) based on dynamically produced execution results.
[EXTERNAL_DOWNLOADS]: The evaluation viewer (viewer.html) fetches the SheetJS library from a well-known CDN (cdn.sheetjs.com) to enable the local visualization of spreadsheet data generated during skill testing. This is a standard and neutrally reported dependency.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from skill drafts, test queries, and user feedback (feedback.json), which serves as an ingestion point for potential indirect prompt injection. The workflow lacks explicit input sanitization scripts but incorporates a human-in-the-loop review process as a boundary marker. Capability inventory includes the ability to execute the claude CLI and perform file system operations.

skill-creator