The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses Python's subprocess module to execute system commands. Specifically, scripts/run_eval.py and scripts/improve_description.py invoke the claude CLI tool to perform evaluations and generate optimized descriptions. Additionally, eval-viewer/generate_review.py uses lsof and os.kill to identify and terminate processes listening on specific network ports to manage its local HTTP server.
[INDIRECT_PROMPT_INJECTION]: The skill represents a surface for indirect prompt injection as it ingests untrusted data that is subsequently processed by AI agents.
Ingestion points: User-provided test queries in evals.json and eval_set.json, and human feedback in feedback.json.
Boundary markers: The skill does not implement explicit delimiters or instruction-following guardrails when passing this data to sub-agents.
Capability inventory: The skill has capabilities for shell execution (via subprocess), local file writing (creating .skill packages and workspace iterations), and starting a local HTTP server.
Sanitization: There is no evidence of input validation or sanitization for the queries and feedback before they are interpolated into sub-agent prompts.
[EXTERNAL_DOWNLOADS]: The eval-viewer/viewer.html template includes a script tag that loads the SheetJS (xlsx) library from a Content Delivery Network (cdn.sheetjs.com) to render spreadsheet outputs in the browser. This is a well-known service used for the skill's primary purpose.
[DYNAMIC_EXECUTION]: The skill dynamically generates and writes SKILL.md files to the .claude/commands/ directory in scripts/run_eval.py to test how Claude Code responds to new skill descriptions. It also programmatically packages skill directories into .skill (ZIP) files using scripts/package_skill.py.

skill-creator