The Agent Skills Directory

[COMMAND_EXECUTION]: The skill makes extensive use of the subprocess module to execute system commands and external CLI tools. Specifically:
scripts/run_eval.py uses subprocess.Popen to call the claude CLI for running evaluation prompts.
scripts/improve_description.py uses subprocess.run to call the claude CLI to generate optimized skill descriptions.
eval-viewer/generate_review.py uses subprocess.run to execute the lsof command to manage network ports for its local server.
[DYNAMIC_EXECUTION]: The skill dynamically generates and executes commands based on user-provided data and skill logic. It creates temporary command files in .claude/commands/ to register skills at runtime and executes them via the claude CLI.
[INDIRECT_PROMPT_INJECTION]: The skill acts as a platform for processing untrusted data (test prompts) which are then executed by subagents with high capabilities.
Ingestion points: Evaluation prompts are defined in evals/evals.json and processed in SKILL.md (Step 1) and scripts/run_eval.py.
Boundary markers: The skill uses markdown structures to delineate tasks for subagents, but it lacks rigorous sanitization for the interpolated test prompts.
Capability inventory: The environment has high capabilities, including file system read/write access (via zipfile and Path.write_text) and subprocess execution.
Sanitization: While skill descriptions are escaped into YAML block scalars in run_eval.py, user-provided prompts are passed directly as command-line arguments to the claude CLI.
[EXTERNAL_DOWNLOADS]: The evaluation viewer template (eval-viewer/viewer.html) loads external resources from well-known and trusted services:
Fetches fonts from Google's official Font API (fonts.googleapis.com).
Downloads the SheetJS library from a well-known CDN (cdn.sheetjs.com) for spreadsheet processing. These references are documented as safe per internal policy.

skill-creator