The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses the Python subprocess module to execute the claude CLI tool. This is performed across several scripts (run_eval.py, improve_description.py, run_loop.py) to test skill triggering and generate improved skill descriptions based on model output.
Evidence: subprocess.Popen calls in run_eval.py and subprocess.run calls in improve_description.py and generate_review.py.
[EXTERNAL_DOWNLOADS]: The evaluation results viewer (eval-viewer/viewer.html) loads the SheetJS library from a remote CDN (cdn.sheetjs.com) to enable Excel file rendering in the browser.
Evidence: <script src="https://cdn.sheetjs.com/xlsx-0.20.3/package/dist/xlsx.full.min.js" ...> in eval-viewer/viewer.html.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection because it ingests untrusted data from user-defined eval queries and skill instructions, which are then interpolated into prompts for grader, analyzer, and optimizer subagents.
Ingestion points: evals/evals.json and the SKILL.md file of the skill being developed.
Boundary markers: The skill uses XML-style tags (e.g., <skill_content>) to delimit untrusted data within its internal prompts, which provides some mitigation but does not fully prevent adversarial influence.
Capability inventory: The skill has the capability to write files to the .claude/commands/ directory and execute the claude CLI tool via subprocess.
Sanitization: Content is escaped using html.escape for the HTML reports, but the raw instructional data is passed to the LLM for processing without further structural validation.

skill-creator