The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/run_eval.py executes the claude CLI via subprocess.Popen to perform triggering tests. This is a functional requirement for evaluating how skills are activated by the agent.
[COMMAND_EXECUTION]: In eval-viewer/generate_review.py, the skill uses subprocess.run to call lsof for port management, which is a standard administrative task for local development servers.
[INDIRECT_PROMPT_INJECTION]: The skill processes user-defined evaluation prompts and their outputs through subagents (grader and analyzer), which constitutes a potential indirect prompt injection surface.
Ingestion points: User-provided prompts in evals/evals.json and execution outputs from subagents.
Boundary markers: Content is wrapped in XML-style tags (e.g., <current_description>) to help the agent distinguish instructions from data.
Capability inventory: The skill can execute shell commands and modify local files within the evaluation workspace.
Sanitization: The skill relies on model-level instructions for processing content rather than strict programmatic filtering.
[SAFE]: The packaging utility scripts/package_skill.py uses a dedicated scripts/path_safety.py module to prevent path traversal attacks by validating all filenames and paths before processing them.
[SAFE]: The review server started by eval-viewer/generate_review.py is hardcoded to listen on 127.0.0.1, ensuring that review data and feedback mechanisms are not exposed to the network.

skill-creator