The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/run_eval.py executes the claude CLI via subprocess.Popen to run skill evaluations on test prompts. This is a legitimate use of command execution to interface with the agent's testing environment.\n- [COMMAND_EXECUTION]: The eval-viewer/generate_review.py script uses subprocess.run to call the lsof utility to identify and terminate processes occupying its local server port. This process management is localized to the tool's execution.\n- [EXTERNAL_DOWNLOADS]: The eval-viewer/viewer.html file loads assets from trusted external sources, including Google Fonts and the SheetJS library from a well-known CDN. These downloads are for intended UI functionality.\n- [INDIRECT_PROMPT_INJECTION]: The skill exhibits an indirect prompt injection surface as it ingests user-provided task prompts and intents which are subsequently executed by subagents. While no malicious activity was detected, the skill lacks explicit sanitization for these external inputs.\n
Ingestion points: User-provided eval prompts in evals/evals.json and intent descriptions captured in SKILL.md during the development phase.\n
Boundary markers: None identified in the prompt interpolation logic.\n
Capability inventory: The skill spawns subagents with full agent capabilities to perform the evaluations.\n
Sanitization: No escaping or validation is performed on the test prompts before they are passed to the subagent.

skill-creator