The Agent Skills Directory

[COMMAND_EXECUTION]: Multiple Python scripts, including run_eval.py, improve_description.py, and generate_review.py, use the subprocess module to execute local commands. These scripts manage test evaluations through the claude CLI and control the lifecycle of a local review server using system utilities like lsof and kill. These commands are executed with structured argument lists, which significantly reduces the risk of shell injection.
[PROMPT_INJECTION]: The skill implements an indirect prompt injection surface in its description optimization loop (improve_description.py). Untrusted queries from evals.json are interpolated into prompts for the LLM. To mitigate this risk, the skill uses XML-style tags as boundary markers (e.g., <skill_content>, <scores_summary>) to help the model distinguish between instructions and processed data.
[SAFE]: The skill serves a local HTML viewer (eval-viewer/viewer.html) to display evaluation results to the user. This viewer includes a remote script from SheetJS's CDN (cdn.sheetjs.com) to render Excel files. SheetJS is a well-known service, and the use of its library for local data visualization is considered a safe and standard development practice.
[SAFE]: The skill features a self-optimization loop that updates the SKILL.md frontmatter with improved descriptions generated by an LLM. While this involves dynamic content modification, the process is human-in-the-loop and restricted to the specific skill directory, aligning with the expected functionality of a skill-creation tool.

skill-creator